parsers/renderers: use think from user for nemotron (#13492)

---------

Co-authored-by: Daniel Hiltgen <daniel@ollama.com>

CMakeLists.txt

@@ -54,6 +54,13 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-cp
 
 add_compile_definitions(NDEBUG GGML_VERSION=0x0 GGML_COMMIT=0x0)
 
+# Define GGML version variables for shared library SOVERSION
+# These are required by ggml/src/CMakeLists.txt for proper library versioning
+set(GGML_VERSION_MAJOR 0)
+set(GGML_VERSION_MINOR 0)
+set(GGML_VERSION_PATCH 0)
+set(GGML_VERSION "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
+
 set(GGML_CPU ON)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src)
 set_property(TARGET ggml PROPERTY EXCLUDE_FROM_ALL TRUE)

README.md

@@ -555,7 +555,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Parakeet](https://github.com/parakeet-nest/parakeet) is a GoLang library, made to simplify the development of small generative AI applications with Ollama.
 - [Haverscript](https://github.com/andygill/haverscript) with [examples](https://github.com/andygill/haverscript/tree/main/examples)
 - [Ollama for Swift](https://github.com/mattt/ollama-swift)
-- [Swollama for Swift]([https://github.com/marcusziade/Swollama](https://github.com/guitaripod/Swollama) with [DocC]( https://guitaripod.github.io/Swollama/documentation/swollama)
+- [Swollama for Swift](https://github.com/guitaripod/Swollama) with [DocC](https://guitaripod.github.io/Swollama/documentation/swollama)
 - [GoLamify](https://github.com/prasad89/golamify)
 - [Ollama for Haskell](https://github.com/tusharad/ollama-haskell)
 - [multi-llm-ts](https://github.com/nbonamy/multi-llm-ts) (A Typescript/JavaScript library allowing access to different LLM in a unified API)

api/client.go

@@ -347,7 +347,7 @@ type CreateProgressFunc func(ProgressResponse) error
 // Create creates a model from a [Modelfile]. fn is a progress function that
 // behaves similarly to other methods (see [Client.Pull]).
 //
-// [Modelfile]: https://github.com/ollama/ollama/blob/main/docs/modelfile.md
+// [Modelfile]: https://github.com/ollama/ollama/blob/main/docs/modelfile.mdx
 func (c *Client) Create(ctx context.Context, req *CreateRequest, fn CreateProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/create", req, func(bts []byte) error {
 		var resp ProgressResponse

app/cmd/app/app_darwin.go

@@ -191,13 +191,6 @@ func LaunchNewApp() {
 	C.launchApp(appName)
 }
 
-// Send a request to the main app thread to load a UI page
-func sendUIRequestMessage(path string) {
-	p := C.CString(path)
-	defer C.free(unsafe.Pointer(p))
-	C.uiRequest(p)
-}
-
 func registerLaunchAgent(hasCompletedFirstRun bool) {
 	// Remove any stale Login Item registrations
 	C.unregisterSelfFromLoginItem()

app/cmd/app/app_windows.go

@@ -263,11 +263,6 @@ func createLoginShortcut() error {
 	return nil
 }
 
-// Send a request to the main app thread to load a UI page
-func sendUIRequestMessage(path string) {
-	wintray.SendUIRequestMessage(path)
-}
-
 func LaunchNewApp() {
 }
 

app/dialog/cocoa/dlg.m

@@ -169,37 +169,47 @@ DlgResult fileDlg(FileDlgParams* params) {
 	}
 	
 	NSArray* urls = [panel URLs];
-	if(self->params->allowMultiple && [urls count] >= 1) {
+	if([urls count] == 0) {
+		return DLG_CANCEL;
+	}
+	
+	if(self->params->allowMultiple) {
 		// For multiple files, we need to return all paths separated by null bytes
 		char* bufPtr = self->params->buf;
 		int remainingBuf = self->params->nbuf;
 		
-  // Calculate total required buffer size first
-  int totalSize = 0;
-  for(NSURL* url in urls) {
-      char tempBuf[PATH_MAX];
-      if(![url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX]) {
-          return DLG_URLFAIL;
-      }
-      totalSize += strlen(tempBuf) + 1; // +1 for null terminator
-  }
-  totalSize += 1; // Final null terminator
+		// Calculate total required buffer size first
+		int totalSize = 0;
+		for(NSURL* url in urls) {
+			char tempBuf[PATH_MAX];
+			if(![url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX]) {
+				return DLG_URLFAIL;
+			}
+			totalSize += strlen(tempBuf) + 1; // +1 for null terminator
+		}
+		totalSize += 1; // Final null terminator
 
-  if(totalSize > self->params->nbuf) {
-      // Not enough buffer space
-      return DLG_URLFAIL;
-  }
+		if(totalSize > self->params->nbuf) {
+			// Not enough buffer space
+			return DLG_URLFAIL;
+		}
 
-  // Now actually copy the paths (we know we have space)
-  bufPtr = self->params->buf;
-  for(NSURL* url in urls) {
-      char tempBuf[PATH_MAX];
-      [url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX];
-      int pathLen = strlen(tempBuf);
-      strcpy(bufPtr, tempBuf);
-      bufPtr += pathLen + 1;
-  }
-  *bufPtr = '\0'; // Final null terminator
+		// Now actually copy the paths (we know we have space)
+		bufPtr = self->params->buf;
+		for(NSURL* url in urls) {
+			char tempBuf[PATH_MAX];
+			[url getFileSystemRepresentation:tempBuf maxLength:PATH_MAX];
+			int pathLen = strlen(tempBuf);
+			strcpy(bufPtr, tempBuf);
+			bufPtr += pathLen + 1;
+		}
+		*bufPtr = '\0'; // Final null terminator
+	} else {
+		// Single file/directory selection - write path to buffer
+		NSURL* url = [urls firstObject];
+		if(![url getFileSystemRepresentation:self->params->buf maxLength:self->params->nbuf]) {
+			return DLG_URLFAIL;
+		}
 	}
 	
 	return DLG_OK;

app/dialog/dlgs_windows.go

@@ -15,7 +15,7 @@ const multiFileBufferSize = w32.MAX_PATH * 10
 type WinDlgError int
 
 func (e WinDlgError) Error() string {
-	return fmt.Sprintf("CommDlgExtendedError: %#x", e)
+	return fmt.Sprintf("CommDlgExtendedError: %#x", int(e))
 }
 
 func err() error {

app/server/server.go

@@ -224,9 +224,7 @@ func (s *Server) cmd(ctx context.Context) (*exec.Cmd, error) {
 		if _, err := os.Stat(settings.Models); err == nil {
 			env["OLLAMA_MODELS"] = settings.Models
 		} else {
-			slog.Warn("models path not accessible, clearing models setting", "path", settings.Models, "err", err)
-			settings.Models = ""
-			s.store.SetSettings(settings)
+			slog.Warn("models path not accessible, using default", "path", settings.Models, "err", err)
 		}
 	}
 	if settings.ContextLength > 0 {

app/ui/ui.go

@@ -12,13 +12,13 @@ import (
 	"log/slog"
 	"net/http"
 	"net/http/httputil"
-	"net/url"
 	"os"
 	"runtime"
 	"runtime/debug"
 	"slices"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/google/uuid"
@@ -117,40 +117,66 @@ func (s *Server) log() *slog.Logger {
 
 // ollamaProxy creates a reverse proxy handler to the Ollama server
 func (s *Server) ollamaProxy() http.Handler {
-	ollamaHost := os.Getenv("OLLAMA_HOST")
-	if ollamaHost == "" {
-		ollamaHost = "http://127.0.0.1:11434"
-	}
+	var (
+		proxy   http.Handler
+		proxyMu sync.Mutex
+	)
 
-	if !strings.HasPrefix(ollamaHost, "http://") && !strings.HasPrefix(ollamaHost, "https://") {
-		ollamaHost = "http://" + ollamaHost
-	}
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		proxyMu.Lock()
+		p := proxy
+		proxyMu.Unlock()
 
-	target, err := url.Parse(ollamaHost)
-	if err != nil {
-		s.log().Error("failed to parse OLLAMA_HOST", "error", err, "host", ollamaHost)
-		return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-			http.Error(w, "failed to configure proxy", http.StatusInternalServerError)
-		})
-	}
+		if p == nil {
+			proxyMu.Lock()
+			if proxy == nil {
+				var err error
+				for i := range 2 {
+					if i > 0 {
+						s.log().Warn("ollama server not ready, retrying", "attempt", i+1)
+						time.Sleep(1 * time.Second)
+					}
 
-	s.log().Info("configuring ollama proxy", "target", target.String())
+					err = WaitForServer(context.Background(), 10*time.Second)
+					if err == nil {
+						break
+					}
+				}
 
-	proxy := httputil.NewSingleHostReverseProxy(target)
+				if err != nil {
+					proxyMu.Unlock()
+					s.log().Error("ollama server not ready after retries", "error", err)
+					http.Error(w, "Ollama server is not ready", http.StatusServiceUnavailable)
+					return
+				}
 
-	originalDirector := proxy.Director
-	proxy.Director = func(req *http.Request) {
-		originalDirector(req)
-		req.Host = target.Host
-		s.log().Debug("proxying request", "method", req.Method, "path", req.URL.Path, "target", target.Host)
-	}
+				target := envconfig.Host()
+				s.log().Info("configuring ollama proxy", "target", target.String())
 
-	proxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
-		s.log().Error("proxy error", "error", err, "path", r.URL.Path, "target", target.String())
-		http.Error(w, "proxy error: "+err.Error(), http.StatusBadGateway)
-	}
+				newProxy := httputil.NewSingleHostReverseProxy(target)
 
-	return proxy
+				originalDirector := newProxy.Director
+				newProxy.Director = func(req *http.Request) {
+					originalDirector(req)
+					req.Host = target.Host
+					s.log().Debug("proxying request", "method", req.Method, "path", req.URL.Path, "target", target.Host)
+				}
+
+				newProxy.ErrorHandler = func(w http.ResponseWriter, r *http.Request, err error) {
+					s.log().Error("proxy error", "error", err, "path", r.URL.Path, "target", target.String())
+					http.Error(w, "proxy error: "+err.Error(), http.StatusBadGateway)
+				}
+
+				proxy = newProxy
+				p = newProxy
+			} else {
+				p = proxy
+			}
+			proxyMu.Unlock()
+		}
+
+		p.ServeHTTP(w, r)
+	})
 }
 
 type errHandlerFunc func(http.ResponseWriter, *http.Request) error

app/wintray/eventloop.go

@@ -158,16 +158,16 @@ func (t *winTray) wndProc(hWnd windows.Handle, message uint32, wParam, lParam ui
 	case uint32(UI_REQUEST_MSG_ID):
 		// Requests for the UI must always come from the main event thread
 		l := int(wParam)
-		path := unsafe.String((*byte)(unsafe.Pointer(lParam)), l)
+		path := unsafe.String((*byte)(unsafe.Pointer(lParam)), l) //nolint:govet,gosec
 		t.app.UIRun(path)
 	case WM_COPYDATA:
 		// Handle URL scheme requests from other instances
 		if lParam != 0 {
-			cds := (*COPYDATASTRUCT)(unsafe.Pointer(lParam))
-			if cds.DwData == 1 { // Our identifier for URL scheme messages
+			cds := (*COPYDATASTRUCT)(unsafe.Pointer(lParam)) //nolint:govet,gosec
+			if cds.DwData == 1 {                             // Our identifier for URL scheme messages
 				// Convert the data back to string
 				data := make([]byte, cds.CbData)
-				copy(data, (*[1 << 30]byte)(unsafe.Pointer(cds.LpData))[:cds.CbData:cds.CbData])
+				copy(data, (*[1 << 30]byte)(unsafe.Pointer(cds.LpData))[:cds.CbData:cds.CbData]) //nolint:govet,gosec
 				urlScheme := string(data)
 				handleURLSchemeRequest(urlScheme)
 				lResult = 1 // Return non-zero to indicate success

convert/convert.go

@@ -182,6 +182,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &llama4Model{}
 	case "Mistral3ForConditionalGeneration":
 		conv = &mistral3Model{}
+	case "Ministral3ForCausalLM":
+		conv = &mistral3CausalModel{}
 	case "MixtralForCausalLM":
 		conv = &mixtralModel{}
 	case "GemmaForCausalLM":
@@ -200,6 +202,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &qwen25VLModel{}
 	case "Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration":
 		conv = &qwen3VLModel{}
+	case "Olmo3ForCausalLM":
+		conv = &olmoModel{}
 	case "BertModel":
 		conv = &bertModel{}
 	case "NomicBertModel", "NomicBertMoEModel":

convert/convert_mistral.go

@@ -30,13 +30,15 @@ type mistral3Model struct {
 		HiddenAct             string  `json:"hidden_act"`
 		VocabSize             uint32  `json:"vocab_size"`
 		RopeParameters        struct {
-			BetaFast                  float32 `json:"beta_fast"`
-			BetaSlow                  float32 `json:"beta_slow"`
-			Factor                    float32 `json:"factor"`
-			ScalingBeta               float32 `json:"llama_4_scaling_beta"`
-			OrigMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
-			RopeType                  string  `json:"rope_type"`
-			RopeTheta                 float32 `json:"rope_theta"`
+			BetaFast                  float32  `json:"beta_fast"`
+			BetaSlow                  float32  `json:"beta_slow"`
+			Factor                    float32  `json:"factor"`
+			Llama4ScalingBeta         *float32 `json:"llama_4_scaling_beta"`
+			OrigMaxPositionEmbeddings uint32   `json:"original_max_position_embeddings"`
+			RopeType                  string   `json:"rope_type"`
+			RopeTheta                 float32  `json:"rope_theta"`
+			Mscale                    *float32 `json:"mscale"`
+			MscaleAllDim              *float32 `json:"mscale_all_dim"`
 		} `json:"rope_parameters"`
 	} `json:"text_config"`
 	VisionModel struct {
@@ -50,6 +52,9 @@ type mistral3Model struct {
 		HeadDim           uint32  `json:"head_dim"`
 		HiddenAct         string  `json:"hidden_act"`
 		RopeTheta         float32 `json:"rope_theta"`
+		RopeParameters    struct {
+			RopeTheta float32 `json:"rope_theta"`
+		} `json:"rope_parameters"`
 	} `json:"vision_config"`
 	MultiModalProjectorBias bool   `json:"multimodal_projector_bias"`
 	ProjectorHiddenAct      string `json:"projector_hidden_act"`
@@ -72,10 +77,22 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
 	kv["mistral3.rope.dimension_count"] = cmp.Or(p.TextModel.HeadDim, p.TextModel.HiddenSize/p.TextModel.NumAttentionHeads)
 	kv["mistral3.rope.freq_base"] = cmp.Or(p.TextModel.RopeTheta, p.TextModel.RopeParameters.RopeTheta)
+	kv["mistral3.rope.scaling.factor"] = p.TextModel.RopeParameters.Factor
+	kv["mistral3.rope.scaling.type"] = p.TextModel.RopeParameters.RopeType
+	kv["mistral3.rope.scaling.beta_fast"] = p.TextModel.RopeParameters.BetaFast
+	kv["mistral3.rope.scaling.beta_slow"] = p.TextModel.RopeParameters.BetaSlow
 
+	if p.TextModel.RopeParameters.Mscale != nil {
+		kv["mistral3.rope.scaling.mscale"] = *p.TextModel.RopeParameters.Mscale
+	}
+	if p.TextModel.RopeParameters.MscaleAllDim != nil {
+		kv["mistral3.rope.scaling.mscale_all_dim"] = *p.TextModel.RopeParameters.MscaleAllDim
+	}
 	if p.TextModel.RopeParameters.OrigMaxPositionEmbeddings > 0 {
 		kv["mistral3.rope.scaling.original_context_length"] = p.TextModel.RopeParameters.OrigMaxPositionEmbeddings
-		kv["mistral3.rope.scaling_beta"] = p.TextModel.RopeParameters.ScalingBeta
+	}
+	if p.TextModel.RopeParameters.Llama4ScalingBeta != nil {
+		kv["mistral3.rope.scaling_beta"] = *p.TextModel.RopeParameters.Llama4ScalingBeta
 	}
 
 	// Vision configuration
@@ -88,7 +105,7 @@ func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
 	kv["mistral3.vision.patch_size"] = p.VisionModel.PatchSize
 	kv["mistral3.vision.num_channels"] = p.VisionModel.NumChannels
 	// kv["mistral3.vision.attention.layer_norm_epsilon"] = 1e-05 // Default value
-	kv["mistral3.vision.rope.freq_base"] = p.VisionModel.RopeTheta
+	kv["mistral3.vision.rope.freq_base"] = cmp.Or(p.VisionModel.RopeTheta, p.VisionModel.RopeParameters.RopeTheta)
 
 	// Multimodal configuration
 	kv["mistral3.image_token_index"] = p.ImageTokenIndex

convert/convert_mistral_causal.go

@@ -0,0 +1,181 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type mistral3CausalModel struct {
+	ModelParameters
+
+	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RopeTheta             float32 `json:"rope_theta"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+	HeadDim               uint32  `json:"head_dim"`
+	SlidingWindow         *uint32 `json:"sliding_window"`
+	HiddenAct             string  `json:"hidden_act"`
+	VocabSize             uint32  `json:"vocab_size"`
+	RopeParameters        struct {
+		BetaFast                  float32  `json:"beta_fast"`
+		BetaSlow                  float32  `json:"beta_slow"`
+		Factor                    float32  `json:"factor"`
+		Llama4ScalingBeta         *float32 `json:"llama_4_scaling_beta"`
+		OrigMaxPositionEmbeddings uint32   `json:"original_max_position_embeddings"`
+		RopeType                  string   `json:"rope_type"`
+		RopeTheta                 float32  `json:"rope_theta"`
+		Mscale                    *float32 `json:"mscale"`
+		MscaleAllDim              *float32 `json:"mscale_all_dim"`
+	} `json:"rope_parameters"`
+}
+
+func (p *mistral3CausalModel) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "mistral3"
+	kv["mistral3.vocab_size"] = p.VocabSize
+
+	// Text configuration
+	kv["mistral3.block_count"] = p.NumHiddenLayers
+	kv["mistral3.context_length"] = p.MaxPositionEmbeddings
+	kv["mistral3.embedding_length"] = p.HiddenSize
+	kv["mistral3.feed_forward_length"] = p.IntermediateSize
+	kv["mistral3.attention.head_count"] = p.NumAttentionHeads
+	kv["mistral3.attention.head_count_kv"] = p.NumKeyValueHeads
+	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["mistral3.attention.key_length"] = p.HeadDim
+	kv["mistral3.attention.value_length"] = p.HeadDim
+	kv["mistral3.rope.dimension_count"] = cmp.Or(p.HeadDim, p.HiddenSize/p.NumAttentionHeads)
+	kv["mistral3.rope.freq_base"] = cmp.Or(p.RopeTheta, p.RopeParameters.RopeTheta)
+	kv["mistral3.rope.scaling.factor"] = p.RopeParameters.Factor
+	kv["mistral3.rope.scaling.type"] = p.RopeParameters.RopeType
+	kv["mistral3.rope.scaling.beta_fast"] = p.RopeParameters.BetaFast
+	kv["mistral3.rope.scaling.beta_slow"] = p.RopeParameters.BetaSlow
+
+	if p.RopeParameters.Mscale != nil {
+		kv["mistral3.rope.scaling.mscale"] = *p.RopeParameters.Mscale
+	}
+
+	if p.RopeParameters.MscaleAllDim != nil {
+		kv["mistral3.rope.scaling.mscale_all_dim"] = *p.RopeParameters.MscaleAllDim
+	}
+
+	if p.RopeParameters.OrigMaxPositionEmbeddings > 0 {
+		kv["mistral3.rope.scaling.original_context_length"] = p.RopeParameters.OrigMaxPositionEmbeddings
+		kv["mistral3.rope.scaling_beta"] = *p.RopeParameters.Llama4ScalingBeta
+	}
+
+	if p.RopeParameters.Llama4ScalingBeta != nil {
+		kv["mistral3.rope.scaling_beta"] = *p.RopeParameters.Llama4ScalingBeta
+	}
+
+	return kv
+}
+
+func (p *mistral3CausalModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	for _, t := range ts {
+		if !strings.HasPrefix(t.Name(), "v.") {
+			if strings.HasSuffix(t.Name(), ".attn_q.weight") ||
+				strings.HasSuffix(t.Name(), ".attn_k.weight") {
+				t.SetRepacker(p.repack)
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *mistral3CausalModel) Replacements() []string {
+	return []string{
+		"model.norm", "output_norm",
+		"model.", "",
+		"layers", "blk",
+		"transformer.layers", "blk",
+		"vision_tower", "v",
+		"ln_pre", "encoder_norm",
+		"input_layernorm", "attn_norm",
+		"post_attention_layernorm", "ffn_norm",
+		"embed_tokens", "token_embd",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"attention.q_proj", "attn_q",
+		"attention.k_proj", "attn_k",
+		"attention.v_proj", "attn_v",
+		"attention.o_proj", "attn_output",
+		"attention_norm", "attn_norm",
+		"feed_forward.gate_proj", "ffn_gate",
+		"feed_forward.down_proj", "ffn_down",
+		"feed_forward.up_proj", "ffn_up",
+		"multi_modal_projector", "mm",
+		"ffn_norm", "ffn_norm",
+		"lm_head", "output",
+	}
+}
+
+func (p *mistral3CausalModel) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	var dims []int
+	for _, dim := range shape {
+		dims = append(dims, int(dim))
+	}
+
+	var heads uint32
+	if strings.HasSuffix(name, ".attn_q.weight") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, ".attn_k.weight") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	} else {
+		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}

convert/convert_olmo.go

@@ -0,0 +1,117 @@
+package convert
+
+import (
+	"cmp"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type ropeScaling struct {
+	Factor                    float32 `json:"factor"`
+	OriginalMaxPositionEmbeds uint32  `json:"original_max_position_embeddings"`
+	AttentionFactor           float32 `json:"attention_factor"`
+	BetaFast                  float32 `json:"beta_fast"`
+	BetaSlow                  float32 `json:"beta_slow"`
+	RopeType                  string  `json:"rope_type"`
+	ExtrapolationFactor       float32 `json:"extrapolation_factor"`
+}
+
+type olmoModel struct {
+	ModelParameters
+
+	HiddenSize            uint32       `json:"hidden_size"`
+	NumHiddenLayers       uint32       `json:"num_hidden_layers"`
+	IntermediateSize      uint32       `json:"intermediate_size"`
+	NumAttentionHeads     uint32       `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32       `json:"num_key_value_heads"`
+	MaxPositionEmbeddings uint32       `json:"max_position_embeddings"`
+	RMSNormEPS            float32      `json:"rms_norm_eps"`
+	RopeTheta             float32      `json:"rope_theta"`
+	RopeScaling           *ropeScaling `json:"rope_scaling"`
+	SlidingWindow         uint32       `json:"sliding_window"`
+	LayerTypes            []string     `json:"layer_types"`
+}
+
+var _ ModelConverter = (*olmoModel)(nil)
+
+func (p *olmoModel) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "olmo3"
+	kv["olmo3.block_count"] = p.NumHiddenLayers
+	kv["olmo3.context_length"] = p.MaxPositionEmbeddings
+	kv["olmo3.embedding_length"] = p.HiddenSize
+	kv["olmo3.feed_forward_length"] = p.IntermediateSize
+	kv["olmo3.attention.head_count"] = p.NumAttentionHeads
+	kv["olmo3.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+
+	if p.RopeTheta > 0 {
+		kv["olmo3.rope.freq_base"] = p.RopeTheta
+	}
+
+	if p.RopeScaling != nil {
+		if p.RopeScaling.Factor > 0 {
+			kv["olmo3.rope.scaling.factor"] = p.RopeScaling.Factor
+		}
+		if p.RopeScaling.OriginalMaxPositionEmbeds > 0 {
+			kv["olmo3.rope.scaling.original_context_length"] = p.RopeScaling.OriginalMaxPositionEmbeds
+		}
+		if p.RopeScaling.AttentionFactor > 0 {
+			kv["olmo3.rope.scaling.attn_factor"] = p.RopeScaling.AttentionFactor
+		}
+		if p.RopeScaling.RopeType != "" {
+			kv["olmo3.rope.scaling.type"] = p.RopeScaling.RopeType
+		}
+	}
+
+	if p.RMSNormEPS > 0 {
+		kv["olmo3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	}
+
+	if p.SlidingWindow > 0 {
+		kv["olmo3.attention.sliding_window"] = p.SlidingWindow
+	}
+
+	if len(p.LayerTypes) > 0 {
+		slidingPattern := make([]bool, len(p.LayerTypes))
+		for i, layerType := range p.LayerTypes {
+			slidingPattern[i] = (layerType == "sliding_attention")
+		}
+		kv["olmo3.attention.sliding_window_pattern"] = slidingPattern
+	}
+
+	return kv
+}
+
+func (p *olmoModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	out := make([]*ggml.Tensor, 0, len(ts))
+	for _, t := range ts {
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *olmoModel) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.layers", "blk",
+		"model.norm", "output_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_norm", "attn_k_norm",
+		"post_attention_layernorm", "post_attention_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+	}
+}

convert/reader_safetensors.go

@@ -37,10 +37,6 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
 			return nil, err
 		}
 
-		if n <= 0 || n > 100<<20 {
-			return nil, fmt.Errorf("invalid safetensors file %q (header size: %d): file may be corrupted or a Git LFS pointer", p, n)
-		}
-
 		b := bytes.NewBuffer(make([]byte, 0, n))
 		if _, err = io.CopyN(b, f, n); err != nil {
 			return nil, err

docs/api.md

@@ -50,7 +50,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 Advanced parameters (optional):
 
 - `format`: the format to return a response in. Format can be `json` or a JSON schema
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
 - `system`: system message to (overrides what is defined in the `Modelfile`)
 - `template`: the prompt template to use (overrides what is defined in the `Modelfile`)
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
@@ -507,7 +507,7 @@ The `message` object has the following fields:
 Advanced parameters (optional):
 
 - `format`: the format to return a response in. Format can be `json` or a JSON schema.
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
@@ -1189,7 +1189,7 @@ If you are creating a model from a safetensors directory or from a GGUF file, yo
 - `template`: (optional) the prompt template for the model
 - `license`: (optional) a string or list of strings containing the license or licenses for the model
 - `system`: (optional) a string containing the system prompt for the model
-- `parameters`: (optional) a dictionary of parameters for the model (see [Modelfile](./modelfile.md#valid-parameters-and-values) for a list of parameters)
+- `parameters`: (optional) a dictionary of parameters for the model (see [Modelfile](./modelfile.mdx#valid-parameters-and-values) for a list of parameters)
 - `messages`: (optional) a list of message objects used to create a conversation
 - `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects
 - `quantize` (optional): quantize a non-quantized (e.g. float16) model
@@ -1698,7 +1698,7 @@ Generate embeddings from a model
 Advanced parameters:
 
 - `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 - `dimensions`: number of dimensions for the embedding
 
@@ -1817,7 +1817,7 @@ Generate embeddings from a model
 
 Advanced parameters:
 
-- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.mdx#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
 ### Examples

docs/tools/extract-examples/README.md

@@ -0,0 +1,46 @@
+# extract-examples
+
+Extracts code examples from MDX files to a temp directory so you can run them.
+
+## Usage
+
+```shell
+go run docs/tools/extract-examples/main.go <mdx-file>
+```
+
+## Example
+
+```shell
+go run docs/tools/extract-examples/main.go docs/api/openai-compatibility.mdx
+```
+
+Output:
+
+```
+Extracting code examples to: /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
+
+  - 01_basic.py
+  - 01_basic.js
+  - 01_basic.sh
+  - 02_responses.py
+  - 02_responses.js
+  - 02_responses.sh
+  - 03_vision.py
+  - 03_vision.js
+  - 03_vision.sh
+
+Extracted 9 file(s) to /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
+
+To run examples:
+
+  cd /var/folders/vq/wfm2g6k917d3ldzpjdxc8ph00000gn/T/mdx-examples-3271754368
+  npm install   # for JS examples
+
+then run individual files with `node file.js`, `python file.py`, `bash file.sh`
+```
+
+## How it works
+
+- Parses MDX files looking for fenced code blocks with filenames (e.g., ` ```python basic.py `)
+- Groups examples by their `<CodeGroup>` and prefixes filenames with `01_`, `02_`, etc.
+- Writes all extracted files to a temp directory

docs/tools/extract-examples/main.go

@@ -0,0 +1,137 @@
+package main
+
+import (
+	"bufio"
+	"fmt"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+)
+
+func main() {
+	if len(os.Args) < 2 {
+		fmt.Fprintln(os.Stderr, "Usage: go run extract-examples.go <mdx-file>")
+		os.Exit(1)
+	}
+
+	mdxFile := os.Args[1]
+
+	f, err := os.Open(mdxFile)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Error: %v\n", err)
+		os.Exit(1)
+	}
+	defer f.Close()
+
+	// Create temp directory
+	tempDir, err := os.MkdirTemp("", "mdx-examples-*")
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Error creating temp dir: %v\n", err)
+		os.Exit(1)
+	}
+
+	fmt.Printf("Extracting code examples to: %s\n\n", tempDir)
+
+	// Patterns
+	codeBlockStart := regexp.MustCompile("^```([a-zA-Z0-9_-]+)\\s+([^\\s]+)$")
+	codeGroupStart := regexp.MustCompile("^<CodeGroup")
+	codeGroupEnd := regexp.MustCompile("^</CodeGroup>")
+
+	scanner := bufio.NewScanner(f)
+	inCodeBlock := false
+	inCodeGroup := false
+	var currentFile string
+	var content strings.Builder
+	count := 0
+	codeGroupNum := 0
+
+	for scanner.Scan() {
+		line := scanner.Text()
+
+		// Track CodeGroup boundaries
+		if codeGroupStart.MatchString(line) {
+			inCodeGroup = true
+			codeGroupNum++
+			continue
+		}
+		if codeGroupEnd.MatchString(line) {
+			inCodeGroup = false
+			continue
+		}
+
+		if inCodeBlock {
+			if line == "```" {
+				// End of code block - write file
+				if currentFile != "" {
+					outPath := filepath.Join(tempDir, currentFile)
+					if err := os.WriteFile(outPath, []byte(content.String()), 0o644); err != nil {
+						fmt.Fprintf(os.Stderr, "Error writing %s: %v\n", currentFile, err)
+					} else {
+						fmt.Printf("  - %s\n", currentFile)
+						count++
+					}
+				}
+				inCodeBlock = false
+				currentFile = ""
+				content.Reset()
+			} else {
+				content.WriteString(line)
+				content.WriteString("\n")
+			}
+		} else {
+			if matches := codeBlockStart.FindStringSubmatch(line); matches != nil {
+				inCodeBlock = true
+				filename := matches[2]
+				// Prefix with CodeGroup number if inside a CodeGroup
+				if inCodeGroup {
+					currentFile = fmt.Sprintf("%02d_%s", codeGroupNum, filename)
+				} else {
+					currentFile = filename
+				}
+				content.Reset()
+			}
+		}
+	}
+
+	if err := scanner.Err(); err != nil {
+		fmt.Fprintf(os.Stderr, "Error reading file: %v\n", err)
+		os.Exit(1)
+	}
+
+	// Write package.json for JavaScript dependencies
+	packageJSON := `{
+  "name": "mdx-examples",
+  "type": "module",
+  "dependencies": {
+    "openai": "^4",
+    "ollama": "^0.5"
+  }
+}
+`
+	if err := os.WriteFile(filepath.Join(tempDir, "package.json"), []byte(packageJSON), 0o644); err != nil {
+		fmt.Fprintf(os.Stderr, "Error writing package.json: %v\n", err)
+	}
+
+	// Write pyproject.toml for Python dependencies
+	pyprojectTOML := `[project]
+name = "mdx-examples"
+version = "0.0.0"
+dependencies = [
+    "openai",
+    "ollama",
+]
+`
+	if err := os.WriteFile(filepath.Join(tempDir, "pyproject.toml"), []byte(pyprojectTOML), 0o644); err != nil {
+		fmt.Fprintf(os.Stderr, "Error writing pyproject.toml: %v\n", err)
+	}
+
+	fmt.Printf("\n")
+	fmt.Printf("Extracted %d file(s) to %s\n", count, tempDir)
+	fmt.Printf("\n")
+	fmt.Printf("To run examples:\n")
+	fmt.Printf("\n")
+	fmt.Printf("  cd %s\n  npm install   # for JS examples\n", tempDir)
+	fmt.Printf("\n")
+	fmt.Printf("then run individual files with `node file.js`, `python file.py`, `bash file.sh`\n")
+}

fs/ggml/ggml.go

@@ -13,6 +13,7 @@ import (
 
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/util/bufioutil"
+	"github.com/ollama/ollama/ml"
 )
 
 type GGML struct {
@@ -252,6 +253,7 @@ func (kv KV) OllamaEngineRequired() bool {
 		"deepseekocr",
 		"deepseek2",
 		"nomic-bert",
+		"olmo3",
 	}, kv.Architecture())
 }
 
@@ -550,7 +552,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	}, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention bool) (kv []uint64, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention ml.FlashAttentionType) (kv []uint64, partialOffload, fullOffload uint64) {
 	context *= uint64(numParallel)
 
 	embedding := f.KV().EmbeddingLength()
@@ -791,7 +793,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 		}
 
 		partialOffload = 2 * f.KV().HeadCountMax() / cmp.Or(f.KV().HeadCountKVMin(), 1) * kvTotal / 6
-		if useFlashAttention {
+		if useFlashAttention == ml.FlashAttentionEnabled {
 			// rough estimate of graph size with flash attention on
 			partialOffload = (4*uint64(numParallel) + context>>10 + 110) * format.MebiByte
 		}
@@ -809,6 +811,14 @@ func (f GGML) SupportsKVCacheType(cacheType string) bool {
 	return slices.Contains([]string{"q8_0", "q4_0"}, cacheType)
 }
 
+// KVCacheTypeIsQuantized checks if the requested cache type is a quantized type
+func (f GGML) KVCacheTypeIsQuantized(cacheType string) bool {
+	if cacheType == "" || cacheType == "f16" || cacheType == "f32" || cacheType == "bf16" {
+		return false
+	}
+	return true
+}
+
 // SupportsFlashAttention checks if the model supports flash attention
 func (f GGML) SupportsFlashAttention() bool {
 	_, isEmbedding := f.KV()[fmt.Sprintf("%s.pooling_type", f.KV().Architecture())]
@@ -832,6 +842,7 @@ func (f GGML) FlashAttention() bool {
 		"gemma3",
 		"gptoss", "gpt-oss",
 		"mistral3",
+		"olmo3",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
 	}, f.KV().String("general.architecture"))

integration/embed_test.go

@@ -487,6 +487,63 @@ func TestEmbedTruncation(t *testing.T) {
 	}
 }
 
+// TestEmbedLargeInput tests that embedding models can handle large inputs that would exceed typical batch sizes.
+func TestEmbedLargeInput(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	for _, model := range libraryEmbedModels {
+		model := model
+		t.Run(model, func(t *testing.T) {
+			mctx, mcancel := context.WithTimeout(ctx, 2*time.Minute)
+			defer mcancel()
+
+			// Test with progressively larger inputs
+			testCases := []struct {
+				name       string
+				inputWords int
+			}{
+				{"medium_input_256_words", 256},
+				{"large_input_512_words", 512},
+				{"very_large_input_800_words", 800},
+			}
+
+			for _, tc := range testCases {
+				t.Run(tc.name, func(t *testing.T) {
+					words := make([]string, tc.inputWords)
+					for i := range words {
+						words[i] = "word"
+					}
+					input := strings.Join(words, " ")
+
+					req := api.EmbedRequest{
+						Model:     model,
+						Input:     input,
+						KeepAlive: &api.Duration{Duration: 30 * time.Second},
+					}
+
+					res, err := embedTestHelper(mctx, client, t, req)
+					if err != nil {
+						t.Fatalf("embedding failed for %d words: %v", tc.inputWords, err)
+					}
+
+					if len(res.Embeddings) != 1 {
+						t.Fatalf("expected 1 embedding, got %d", len(res.Embeddings))
+					}
+
+					if len(res.Embeddings[0]) == 0 {
+						t.Fatal("expected non-empty embedding")
+					}
+
+					t.Logf("Successfully embedded %d words (%d tokens)", tc.inputWords, res.PromptEvalCount)
+				})
+			}
+		})
+	}
+}
+
 // TestEmbedStatusCode tests that errors from the embedding endpoint
 // properly preserve their HTTP status codes when returned to the client.
 // This test specifically checks the error handling path in EmbedHandler

llama/llama.cpp/src/llama-arch.cpp

@@ -75,6 +75,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_JAIS,             "jais"             },
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
     { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
+    { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_EXAONE4,          "exaone4"          },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
@@ -1765,6 +1766,39 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_NEMOTRON_H_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            // mamba(2) ssm layers
+            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
+            // attention layers
+            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
+            // dense FFN
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+            // MoE FFN (for MoE layers)
+            { LLM_TENSOR_FFN_GATE_INP,   "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_UP_EXPS,    "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,  "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,"blk.%d.exp_probs_b" },
+            // MoE shared expert layer
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_EXAONE,
         {
@@ -2838,6 +2872,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
         case LLM_ARCH_QWEN3NEXT:
             return true;
         default:

llama/llama.cpp/src/llama-arch.h

@@ -79,6 +79,7 @@ enum llm_arch {
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_NEMOTRON_H,
+    LLM_ARCH_NEMOTRON_H_MOE,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,

llama/llama.cpp/src/llama-graph.cpp

@@ -1089,6 +1089,16 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cur = ggml_relu(ctx0, cur);
                 cb(cur, "ffn_moe_relu", il);
             } break;
+        case LLM_FFN_RELU_SQR:
+            if (gate_exps) {
+                // TODO: add support for gated squared relu
+                GGML_ABORT("fatal error: gated squared relu not implemented");
+            } else {
+                cur = ggml_relu(ctx0, cur);
+                cur = ggml_sqr(ctx0, cur);
+                cb(cur, "ffn_moe_relu_sqr", il);
+            }
+            break;
         default:
             GGML_ABORT("fatal error");
     }

llama/llama.cpp/src/llama-model.cpp

@@ -120,6 +120,8 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
+        case LLM_TYPE_80B_A3B:       return "80B.A3B";
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
         case LLM_TYPE_230B_A10B:     return "230B.A10B";
@@ -1788,6 +1790,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
                 ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
@@ -1803,7 +1806,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+
                 switch (hparams.n_layer) {
+                    case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
                     case 56: type = LLM_TYPE_9B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
@@ -5175,6 +5185,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_NEMOTRON_H:
+            case LLM_ARCH_NEMOTRON_H_MOE:
                 {
                     // mamba2 Mixer SSM params
                     // NOTE: int64_t for tensor dimensions
@@ -5185,6 +5196,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_group    = hparams.ssm_n_group;
                     const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
 
+                    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+                    const int64_t n_ff_shexp = hparams.n_ff_shexp;
+
                     // embeddings
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -5234,12 +5248,26 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias",   i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias",   i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd},         TENSOR_NOT_REQUIRED);
-                        } else {
-                            // mlp layers
-                            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
-                            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, 0);
-                            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
-                            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
+                        }  else {
+                            if (n_expert != 0) {
+                                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
+                                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
+
+                                // MoE branch
+                                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                                // Shared expert branch
+                                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+                                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+
+                            } else {
+                                // mlp layers
+                                layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
+                                layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, 0);
+                                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
+                                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
+                            }
                         }
                     }
                 } break;
@@ -6870,7 +6898,8 @@ void llama_model::print_info() const {
         arch == LLM_ARCH_PLAMO2 ||
         arch == LLM_ARCH_GRANITE_HYBRID ||
         arch == LLM_ARCH_QWEN3NEXT ||
-        arch == LLM_ARCH_NEMOTRON_H) {
+        arch == LLM_ARCH_NEMOTRON_H ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
@@ -6926,7 +6955,8 @@ void llama_model::print_info() const {
     if (arch == LLM_ARCH_MINICPM ||
         arch == LLM_ARCH_GRANITE ||
         arch == LLM_ARCH_GRANITE_MOE ||
-        arch == LLM_ARCH_GRANITE_HYBRID) {
+        arch == LLM_ARCH_GRANITE_HYBRID ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@@ -7107,7 +7137,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     if (arch == LLM_ARCH_FALCON_H1) {
                         filter_attn = [&](int32_t) { return true; };
                         filter_recr = [&](int32_t) { return true; };
-                    } else if (arch == LLM_ARCH_NEMOTRON_H) {
+                    } else if (arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) {
                         filter_attn = [&](int32_t il) {
                             return !hparams.is_recurrent(il) && hparams.n_ff(il) == 0;
                         };
@@ -7478,6 +7508,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_nemotron>(*this, params);
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 llm = std::make_unique<llm_build_nemotron_h>(*this, params);
             } break;
@@ -7765,6 +7796,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARWKV7:
         case LLM_ARCH_WAVTOKENIZER_DEC:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             return LLAMA_ROPE_TYPE_NONE;
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values

llama/llama.cpp/src/llama-model.h

@@ -114,6 +114,7 @@ enum llm_type {
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_31B_A3_5B,
     LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air

llama/llama.cpp/src/models/nemotron-h.cpp

@@ -107,12 +107,41 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
 }
 
 ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) {
-    cur = build_ffn(cur,
-            model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
-            NULL, NULL, NULL,
-            model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-            NULL, LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
-    cb(cur, "ffn_out", il);
+    if (model.layers[il].ffn_gate_inp == nullptr) {
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                NULL,                      NULL,                        NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+    } else {
+        ggml_tensor * ffn_inp = cur;
+        ggml_tensor * moe_out =
+            build_moe_ffn(ffn_inp,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    nullptr, // no gate
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
+                    true, hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up_shexp,  NULL, NULL,
+                    NULL /* no gate */           ,  NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+        cb(ffn_shexp, "ffn_shexp", il);
+
+        cur = ggml_add(ctx0, moe_out, ffn_shexp);
+        cb(cur, "ffn_out", il);
+    }
 
     cur = build_cvec(cur, il);
     cb(cur, "l_out", il);

llama/llama.go

@@ -118,18 +118,22 @@ type ContextParams struct {
 	c C.struct_llama_context_params
 }
 
-func NewContextParams(numCtx int, batchSize int, numSeqMax int, threads int, flashAttention bool, kvCacheType string) ContextParams {
+func NewContextParams(numCtx int, batchSize int, numSeqMax int, threads int, flashAttention ml.FlashAttentionType, kvCacheType string) ContextParams {
 	params := C.llama_context_default_params()
 	params.n_ctx = C.uint(numCtx)
-	params.n_batch = C.uint(batchSize)
+	params.n_batch = C.uint(batchSize * numSeqMax)
+	params.n_ubatch = C.uint(batchSize)
 	params.n_seq_max = C.uint(numSeqMax)
 	params.n_threads = C.int(threads)
 	params.n_threads_batch = params.n_threads
 	params.embeddings = C.bool(true)
-	if flashAttention {
-		params.flash_attn_type = C.LLAMA_FLASH_ATTN_TYPE_ENABLED
-	} else {
-		params.flash_attn_type = C.LLAMA_FLASH_ATTN_TYPE_DISABLED
+	switch flashAttention {
+	case ml.FlashAttentionEnabled:
+		params.flash_attn_type = int32(C.LLAMA_FLASH_ATTN_TYPE_ENABLED)
+	case ml.FlashAttentionDisabled:
+		params.flash_attn_type = int32(C.LLAMA_FLASH_ATTN_TYPE_DISABLED)
+	case ml.FlashAttentionAuto:
+		params.flash_attn_type = int32(C.LLAMA_FLASH_ATTN_TYPE_AUTO)
 	}
 	params.type_k = kvCacheTypeFromStr(strings.ToLower(kvCacheType))
 	params.type_v = kvCacheTypeFromStr(strings.ToLower(kvCacheType))

llama/patches/0032-llama-add-support-for-NVIDIA-Nemotron-Nano-3.patch

@@ -0,0 +1,586 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Daniel Bevenius <daniel.bevenius@gmail.com>
+Date: Mon, 15 Dec 2025 15:13:49 +0100
+Subject: [PATCH] llama : add support for NVIDIA Nemotron Nano 3
+
+This commit adds support for the NVIDIA Nemotron Nano 3 model, enabling
+the conversion and running of this model.
+
+fix indentation in llama-graph.cpp
+
+fix indentation and move ffn_inp
+
+convert : fix modify_tensors in NemotronHModel to call super()
+
+fix pyright error
+
+fix flake8 errors
+---
+ convert_hf_to_gguf.py          | 116 +++++++++++++++++++++++++++++++--
+ gguf-py/gguf/constants.py      |  29 +++++++++
+ gguf-py/gguf/tensor_mapping.py |   9 ++-
+ src/llama-arch.cpp             |  35 ++++++++++
+ src/llama-arch.h               |   1 +
+ src/llama-graph.cpp            |  10 +++
+ src/llama-model.cpp            |  50 +++++++++++---
+ src/llama-model.h              |   1 +
+ src/models/nemotron-h.cpp      |  41 ++++++++++--
+ 9 files changed, 269 insertions(+), 23 deletions(-)
+
+diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
+index 867bc9053..57ec2faac 100755
+--- a/convert_hf_to_gguf.py
++++ b/convert_hf_to_gguf.py
+@@ -8601,8 +8601,18 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
+ class NemotronHModel(GraniteHybridModel):
+     """Hybrid mamba2/attention model from NVIDIA"""
+     model_arch = gguf.MODEL_ARCH.NEMOTRON_H
++    is_moe: bool = False
+ 
+     def __init__(self, *args, **kwargs):
++        # We have to determine the correct model architecture (MoE vs non-MoE) before
++        # calling the parent __init__. This is because the parent constructor
++        # uses self.model_arch to build the tensor name map, and all MoE-specific
++        # mappings would be missed if it were called with the default non-MoE arch.
++        hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
++        if "num_experts_per_tok" in hparams:
++            self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
++            self.is_moe = True
++
+         super().__init__(*args, **kwargs)
+ 
+         # Save the top-level head_dim for later
+@@ -8614,9 +8624,11 @@ class NemotronHModel(GraniteHybridModel):
+ 
+         # Update the ssm / attn / mlp layers
+         # M: Mamba2, *: Attention, -: MLP
++        # MoE:
++        # M: Mamba2, *: Attention, E: Expert
+         hybrid_override_pattern = self.hparams["hybrid_override_pattern"]
+         self._ssm_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == "M"]
+-        self._mlp_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == "-"]
++        self._mlp_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == ("E" if self.is_moe else "-")]
+ 
+     def get_attn_layers(self):
+         hybrid_override_pattern = self.hparams["hybrid_override_pattern"]
+@@ -8632,10 +8644,28 @@ class NemotronHModel(GraniteHybridModel):
+         # Set feed_forward_length
+         # NOTE: This will trigger an override warning. This is preferrable to
+         #   duplicating all the parent logic
+-        n_ff = self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"])
+-        self.gguf_writer.add_feed_forward_length([
+-            n_ff if i in self._mlp_layers else 0 for i in range(self.block_count)
+-        ])
++        if not self.is_moe:
++            n_ff = self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"])
++            self.gguf_writer.add_feed_forward_length([
++                n_ff if i in self._mlp_layers else 0 for i in range(self.block_count)
++            ])
++        else:
++            moe_intermediate_size = self.hparams["moe_intermediate_size"]
++            self.gguf_writer.add_feed_forward_length([
++                moe_intermediate_size if i in self._mlp_layers else 0 for i in range(self.block_count)
++            ])
++            self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
++            self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
++            self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["moe_shared_expert_intermediate_size"])
++            self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
++            self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
++            self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
++            self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
++            self.gguf_writer.add_expert_group_count(self.hparams["n_group"])
++
++            # number of experts used per token (top-k)
++            if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
++                self.gguf_writer.add_expert_used_count(n_experts_used)
+ 
+     def set_vocab(self):
+         super().set_vocab()
+@@ -8643,7 +8673,81 @@ class NemotronHModel(GraniteHybridModel):
+         # The tokenizer _does_ add a BOS token (via post_processor type
+         # TemplateProcessing) but does not set add_bos_token to true in the
+         # config, so we need to explicitly override it here.
+-        self.gguf_writer.add_add_bos_token(True)
++        if not self.is_moe:
++            self.gguf_writer.add_add_bos_token(True)
++
++    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
++        if self.is_moe and bid is not None:
++            if name.endswith("mixer.gate.e_score_correction_bias"):
++                new_name = name.replace("e_score_correction_bias", "e_score_correction_bias.bias")
++                mapped_name = self.map_tensor_name(new_name)
++                return [(mapped_name, data_torch)]
++
++            if name.endswith("mixer.dt_bias"):
++                new_name = name.replace("dt_bias", "dt.bias")
++                mapped_name = self.map_tensor_name(new_name)
++                return [(mapped_name, data_torch)]
++
++            if name.endswith("mixer.conv1d.weight"):
++                squeezed_data = data_torch.squeeze()
++                mapped_name = self.map_tensor_name(name)
++                return [(mapped_name, squeezed_data)]
++
++            if name.endswith("mixer.A_log"):
++                transformed_data = -torch.exp(data_torch)
++                reshaped_data = transformed_data.squeeze().reshape(-1, 1)
++                mapped_name = self.map_tensor_name(name)
++                return [(mapped_name, reshaped_data)]
++
++            if name.endswith("mixer.D"):
++                reshaped_data = data_torch.squeeze().reshape(-1, 1)
++                mapped_name = self.map_tensor_name(name)
++                return [(mapped_name, reshaped_data)]
++
++            if name.endswith("mixer.norm.weight"):
++                reshaped_data = data_torch.reshape(8, 512)
++                mapped_name = self.map_tensor_name(name)
++                return [(mapped_name, reshaped_data)]
++
++            if name.find("mixer.experts") != -1:
++                n_experts = self.hparams["n_routed_experts"]
++                assert bid is not None
++
++                if self._experts is None:
++                    self._experts = [{} for _ in range(self.block_count)]
++
++                self._experts[bid][name] = data_torch
++
++                if len(self._experts[bid]) >= n_experts * 2:
++                    # merge the experts into a single tensor
++                    tensors: list[tuple[str, Tensor]] = []
++                    for w_name in ["down_proj", "up_proj"]:
++                        datas: list[Tensor] = []
++
++                        for xid in range(n_experts):
++                            ename = f"backbone.layers.{bid}.mixer.experts.{xid}.{w_name}.weight"
++                            datas.append(self._experts[bid][ename])
++                            del self._experts[bid][ename]
++
++                        data_torch = torch.stack(datas, dim=0)
++                        merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
++                        new_name = self.map_tensor_name(merged_name)
++                        tensors.append((new_name, data_torch))
++
++                    return tensors
++                else:
++                    return []
++
++        return super().modify_tensors(data_torch, name, bid)
++
++    def prepare_tensors(self):
++        super().prepare_tensors()
++
++        if self._experts is not None:
++            # flatten `list[dict[str, Tensor]]` into `list[str]`
++            experts = [k for d in self._experts for k in d.keys()]
++            if len(experts) > 0:
++                raise ValueError(f"Unprocessed experts: {experts}")
+ 
+ 
+ @ModelBase.register("BailingMoeForCausalLM")
+diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
+index 2b8489c59..1852428b4 100644
+--- a/gguf-py/gguf/constants.py
++++ b/gguf-py/gguf/constants.py
+@@ -413,6 +413,7 @@ class MODEL_ARCH(IntEnum):
+     JAIS             = auto()
+     NEMOTRON         = auto()
+     NEMOTRON_H       = auto()
++    NEMOTRON_H_MOE   = auto()
+     EXAONE           = auto()
+     EXAONE4          = auto()
+     GRANITE          = auto()
+@@ -786,6 +787,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
+     MODEL_ARCH.JAIS:             "jais",
+     MODEL_ARCH.NEMOTRON:         "nemotron",
+     MODEL_ARCH.NEMOTRON_H:       "nemotron_h",
++    MODEL_ARCH.NEMOTRON_H_MOE:   "nemotron_h_moe",
+     MODEL_ARCH.EXAONE:           "exaone",
+     MODEL_ARCH.EXAONE4:          "exaone4",
+     MODEL_ARCH.GRANITE:          "granite",
+@@ -2529,6 +2531,33 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
+         MODEL_TENSOR.FFN_DOWN,
+         MODEL_TENSOR.FFN_UP,
+     ],
++    MODEL_ARCH.NEMOTRON_H_MOE: [
++        MODEL_TENSOR.TOKEN_EMBD,
++        MODEL_TENSOR.OUTPUT_NORM,
++        MODEL_TENSOR.OUTPUT,
++        MODEL_TENSOR.ATTN_NORM,
++        MODEL_TENSOR.SSM_IN,
++        MODEL_TENSOR.SSM_CONV1D,
++        MODEL_TENSOR.SSM_DT,
++        MODEL_TENSOR.SSM_A,
++        MODEL_TENSOR.SSM_D,
++        MODEL_TENSOR.SSM_NORM,
++        MODEL_TENSOR.SSM_OUT,
++        MODEL_TENSOR.ATTN_Q,
++        MODEL_TENSOR.ATTN_K,
++        MODEL_TENSOR.ATTN_V,
++        MODEL_TENSOR.ATTN_OUT,
++        MODEL_TENSOR.FFN_DOWN,
++        MODEL_TENSOR.FFN_UP,
++        # experts
++        MODEL_TENSOR.FFN_GATE_INP,
++        MODEL_TENSOR.FFN_UP_EXP,
++        MODEL_TENSOR.FFN_DOWN_EXP,
++        # shared expert
++        MODEL_TENSOR.FFN_DOWN_SHEXP,
++        MODEL_TENSOR.FFN_UP_SHEXP,
++        MODEL_TENSOR.FFN_EXP_PROBS_B,
++    ],
+     MODEL_ARCH.EXAONE: [
+         MODEL_TENSOR.TOKEN_EMBD,
+         MODEL_TENSOR.OUTPUT_NORM,
+diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
+index d9c87da19..7a3c7c5e0 100644
+--- a/gguf-py/gguf/tensor_mapping.py
++++ b/gguf-py/gguf/tensor_mapping.py
+@@ -377,6 +377,7 @@ class TensorNameMap:
+             "model.layers.{bid}.feed_forward.gate",               # lfm2moe
+             "model.layers.{bid}.mlp.router.gate",               # afmoe
+             "layers.{bid}.gate",                                # mistral-large
++            "backbone.layers.{bid}.mixer.gate",                 # nemotron-h-moe
+         ),
+ 
+         MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
+@@ -390,6 +391,7 @@ class TensorNameMap:
+             "model.layers.{bid}.mlp.expert_bias",                           # afmoe
+             "model.layers.{bid}.feed_forward.expert_bias",                  # lfm2moe
+             "model.layers.{bid}.block_sparse_moe.e_score_correction",       # minimax-m2
++            "backbone.layers.{bid}.mixer.gate.e_score_correction_bias"      # nemotron-h-moe
+         ),
+ 
+         # Feed-forward up
+@@ -438,7 +440,7 @@ class TensorNameMap:
+             "layers.{bid}.feed_forward.experts.w3",                 # mixtral (merged)
+             "transformer.decoder_layer.{bid}.moe.linear_v",         # Grok (merged)
+             "transformer.blocks.{bid}.ffn.experts.mlp.v1",          # dbrx
+-            "model.layers.{bid}.mlp.experts.up_proj",               # qwen2moe olmoe (merged) ernie4.5-moe
++            "model.layers.{bid}.mlp.experts.up_proj",               # qwen2moe olmoe (merged) ernie4.5-moe, nemotron-h-moe (merged)
+             "model.layers.{bid}.block_sparse_moe.experts.w3",       # phimoe (merged)
+             "model.layers.{bid}.feed_forward.experts.up_proj",      # llama4
+             "encoder.layers.{bid}.mlp.experts.mlp.w1",              # nomic-bert-moe
+@@ -452,6 +454,7 @@ class TensorNameMap:
+             "model.layers.{bid}.feed_forward.down_proj",
+             "model.layers.{bid}.mlp.shared_mlp.up_proj",             # hunyuan
+             "layers.{bid}.shared_experts.w3",                        # mistral-large
++            "backbone.layers.{bid}.mixer.shared_experts.up_proj",    # nemotron-h-moe
+         ),
+ 
+         MODEL_TENSOR.FFN_UP_CHEXP: (
+@@ -546,7 +549,7 @@ class TensorNameMap:
+             "layers.{bid}.feed_forward.experts.w2",                 # mixtral (merged)
+             "transformer.decoder_layer.{bid}.moe.linear_1",         # Grok (merged)
+             "transformer.blocks.{bid}.ffn.experts.mlp.w2",          # dbrx
+-            "model.layers.{bid}.mlp.experts.down_proj",             # qwen2moe olmoe (merged) ernie4.5-moe
++            "model.layers.{bid}.mlp.experts.down_proj",             # qwen2moe olmoe (merged) ernie4.5-moe nemotron-h-moe (merged)
+             "model.layers.{bid}.block_sparse_moe.output_linear",    # granitemoe
+             "model.layers.{bid}.block_sparse_moe.experts.w2",       # phimoe (merged)
+             "model.layers.{bid}.feed_forward.experts.down_proj",    # llama4
+@@ -561,6 +564,7 @@ class TensorNameMap:
+             "model.layers.{bid}.shared_mlp.output_linear",             # granitemoe
+             "model.layers.{bid}.mlp.shared_mlp.down_proj",             # hunyuan
+             "layers.{bid}.shared_experts.w2",                          # mistral-large
++            "backbone.layers.{bid}.mixer.shared_experts.down_proj",    # nemotron-h-moe
+         ),
+ 
+         MODEL_TENSOR.FFN_DOWN_CHEXP: (
+@@ -704,6 +708,7 @@ class TensorNameMap:
+             "model.layers.{bid}.mamba.dt_proj",         # jamba falcon-h1 granite-hybrid
+             "model.layers.layers.{bid}.mixer.dt_proj",  # plamo2
+             "model.layers.{bid}.linear_attn.dt_proj",   # qwen3next
++            "backbone.layers.{bid}.mixer.dt",           # nemotron-h-moe
+         ),
+ 
+         MODEL_TENSOR.SSM_DT_NORM: (
+diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
+index a5fe4f66c..ac8b5e033 100644
+--- a/src/llama-arch.cpp
++++ b/src/llama-arch.cpp
+@@ -75,6 +75,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+     { LLM_ARCH_JAIS,             "jais"             },
+     { LLM_ARCH_NEMOTRON,         "nemotron"         },
+     { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
++    { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
+     { LLM_ARCH_EXAONE,           "exaone"           },
+     { LLM_ARCH_EXAONE4,          "exaone4"          },
+     { LLM_ARCH_RWKV6,            "rwkv6"            },
+@@ -1765,6 +1766,39 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+             { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+         },
+     },
++    {
++        LLM_ARCH_NEMOTRON_H_MOE,
++        {
++            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
++            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
++            { LLM_TENSOR_OUTPUT,         "output" },
++            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
++            // mamba(2) ssm layers
++            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
++            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
++            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
++            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
++            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
++            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
++            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
++            // attention layers
++            { LLM_TENSOR_ATTN_Q,         "blk.%d.attn_q" },
++            { LLM_TENSOR_ATTN_K,         "blk.%d.attn_k" },
++            { LLM_TENSOR_ATTN_V,         "blk.%d.attn_v" },
++            { LLM_TENSOR_ATTN_OUT,       "blk.%d.attn_output" },
++            // dense FFN
++            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
++            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
++            // MoE FFN (for MoE layers)
++            { LLM_TENSOR_FFN_GATE_INP,   "blk.%d.ffn_gate_inp" },
++            { LLM_TENSOR_FFN_UP_EXPS,    "blk.%d.ffn_up_exps" },
++            { LLM_TENSOR_FFN_DOWN_EXPS,  "blk.%d.ffn_down_exps" },
++            { LLM_TENSOR_FFN_EXP_PROBS_B,"blk.%d.exp_probs_b" },
++            // MoE shared expert layer
++            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
++            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
++        },
++    },
+     {
+         LLM_ARCH_EXAONE,
+         {
+@@ -2838,6 +2872,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
+         case LLM_ARCH_LFM2:
+         case LLM_ARCH_LFM2MOE:
+         case LLM_ARCH_NEMOTRON_H:
++        case LLM_ARCH_NEMOTRON_H_MOE:
+         case LLM_ARCH_QWEN3NEXT:
+             return true;
+         default:
+diff --git a/src/llama-arch.h b/src/llama-arch.h
+index ec9e3a6df..61d73786c 100644
+--- a/src/llama-arch.h
++++ b/src/llama-arch.h
+@@ -79,6 +79,7 @@ enum llm_arch {
+     LLM_ARCH_JAIS,
+     LLM_ARCH_NEMOTRON,
+     LLM_ARCH_NEMOTRON_H,
++    LLM_ARCH_NEMOTRON_H_MOE,
+     LLM_ARCH_EXAONE,
+     LLM_ARCH_EXAONE4,
+     LLM_ARCH_RWKV6,
+diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
+index 43620df78..763202d87 100644
+--- a/src/llama-graph.cpp
++++ b/src/llama-graph.cpp
+@@ -1089,6 +1089,16 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
+                 cur = ggml_relu(ctx0, cur);
+                 cb(cur, "ffn_moe_relu", il);
+             } break;
++        case LLM_FFN_RELU_SQR:
++            if (gate_exps) {
++                // TODO: add support for gated squared relu
++                GGML_ABORT("fatal error: gated squared relu not implemented");
++            } else {
++                cur = ggml_relu(ctx0, cur);
++                cur = ggml_sqr(ctx0, cur);
++                cb(cur, "ffn_moe_relu_sqr", il);
++            }
++            break;
+         default:
+             GGML_ABORT("fatal error");
+     }
+diff --git a/src/llama-model.cpp b/src/llama-model.cpp
+index 3c503b424..94dee78c3 100644
+--- a/src/llama-model.cpp
++++ b/src/llama-model.cpp
+@@ -120,6 +120,8 @@ const char * llm_type_name(llm_type type) {
+         case LLM_TYPE_16B_A1B:       return "16B.A1B";
+         case LLM_TYPE_21B_A3B:       return "21B.A3B";
+         case LLM_TYPE_30B_A3B:       return "30B.A3B";
++        case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
++        case LLM_TYPE_80B_A3B:       return "80B.A3B";
+         case LLM_TYPE_100B_A6B:      return "100B.A6B";
+         case LLM_TYPE_106B_A12B:     return "106B.A12B";
+         case LLM_TYPE_230B_A10B:     return "230B.A10B";
+@@ -1788,6 +1790,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+                 }
+             } break;
+         case LLM_ARCH_NEMOTRON_H:
++        case LLM_ARCH_NEMOTRON_H_MOE:
+             {
+                 ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+                 ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+@@ -1803,7 +1806,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+ 
+                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+ 
++                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
++                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
++                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
++                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
++                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
++
+                 switch (hparams.n_layer) {
++                    case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
+                     case 56: type = LLM_TYPE_9B; break;
+                     default: type = LLM_TYPE_UNKNOWN;
+                 }
+@@ -5175,6 +5185,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+                     }
+                 } break;
+             case LLM_ARCH_NEMOTRON_H:
++            case LLM_ARCH_NEMOTRON_H_MOE:
+                 {
+                     // mamba2 Mixer SSM params
+                     // NOTE: int64_t for tensor dimensions
+@@ -5185,6 +5196,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+                     const int64_t n_group    = hparams.ssm_n_group;
+                     const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
+ 
++                    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
++                    const int64_t n_ff_shexp = hparams.n_ff_shexp;
++
+                     // embeddings
+                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+ 
+@@ -5234,12 +5248,26 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+                             layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias",   i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED);
+                             layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias",   i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED);
+                             layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd},         TENSOR_NOT_REQUIRED);
+-                        } else {
+-                            // mlp layers
+-                            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
+-                            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, 0);
+-                            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
+-                            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
++                        }  else {
++                            if (n_expert != 0) {
++                                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
++                                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
++
++                                // MoE branch
++                                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
++                                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
++
++                                // Shared expert branch
++                                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
++                                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
++
++                            } else {
++                                // mlp layers
++                                layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
++                                layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, 0);
++                                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
++                                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
++                            }
+                         }
+                     }
+                 } break;
+@@ -6870,7 +6898,8 @@ void llama_model::print_info() const {
+         arch == LLM_ARCH_PLAMO2 ||
+         arch == LLM_ARCH_GRANITE_HYBRID ||
+         arch == LLM_ARCH_QWEN3NEXT ||
+-        arch == LLM_ARCH_NEMOTRON_H) {
++        arch == LLM_ARCH_NEMOTRON_H ||
++        arch == LLM_ARCH_NEMOTRON_H_MOE) {
+         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
+         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
+         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
+@@ -6926,7 +6955,8 @@ void llama_model::print_info() const {
+     if (arch == LLM_ARCH_MINICPM ||
+         arch == LLM_ARCH_GRANITE ||
+         arch == LLM_ARCH_GRANITE_MOE ||
+-        arch == LLM_ARCH_GRANITE_HYBRID) {
++        arch == LLM_ARCH_GRANITE_HYBRID ||
++        arch == LLM_ARCH_NEMOTRON_H_MOE) {
+         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
+         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
+         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
+@@ -7107,7 +7137,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
+                     if (arch == LLM_ARCH_FALCON_H1) {
+                         filter_attn = [&](int32_t) { return true; };
+                         filter_recr = [&](int32_t) { return true; };
+-                    } else if (arch == LLM_ARCH_NEMOTRON_H) {
++                    } else if (arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) {
+                         filter_attn = [&](int32_t il) {
+                             return !hparams.is_recurrent(il) && hparams.n_ff(il) == 0;
+                         };
+@@ -7478,6 +7508,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
+                 llm = std::make_unique<llm_build_nemotron>(*this, params);
+             } break;
+         case LLM_ARCH_NEMOTRON_H:
++        case LLM_ARCH_NEMOTRON_H_MOE:
+             {
+                 llm = std::make_unique<llm_build_nemotron_h>(*this, params);
+             } break;
+@@ -7765,6 +7796,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+         case LLM_ARCH_ARWKV7:
+         case LLM_ARCH_WAVTOKENIZER_DEC:
+         case LLM_ARCH_NEMOTRON_H:
++        case LLM_ARCH_NEMOTRON_H_MOE:
+             return LLAMA_ROPE_TYPE_NONE;
+ 
+         // use what we call a normal RoPE, operating on pairs of consecutive head values
+diff --git a/src/llama-model.h b/src/llama-model.h
+index cbf4e1bfa..b378b23ec 100644
+--- a/src/llama-model.h
++++ b/src/llama-model.h
+@@ -114,6 +114,7 @@ enum llm_type {
+     LLM_TYPE_16B_A1B,
+     LLM_TYPE_21B_A3B, // Ernie MoE small
+     LLM_TYPE_30B_A3B,
++    LLM_TYPE_31B_A3_5B,
+     LLM_TYPE_80B_A3B, // Qwen3 Next
+     LLM_TYPE_100B_A6B,
+     LLM_TYPE_106B_A12B, // GLM-4.5-Air
+diff --git a/src/models/nemotron-h.cpp b/src/models/nemotron-h.cpp
+index 541434888..eb135e63f 100644
+--- a/src/models/nemotron-h.cpp
++++ b/src/models/nemotron-h.cpp
+@@ -107,12 +107,41 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
+ }
+ 
+ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) {
+-    cur = build_ffn(cur,
+-            model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+-            NULL, NULL, NULL,
+-            model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+-            NULL, LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
+-    cb(cur, "ffn_out", il);
++    if (model.layers[il].ffn_gate_inp == nullptr) {
++        cur = build_ffn(cur,
++                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
++                NULL,                      NULL,                        NULL,
++                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
++                NULL,
++                LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
++        cb(cur, "ffn_out", il);
++    } else {
++        ggml_tensor * ffn_inp = cur;
++        ggml_tensor * moe_out =
++            build_moe_ffn(ffn_inp,
++                    model.layers[il].ffn_gate_inp,
++                    model.layers[il].ffn_up_exps,
++                    nullptr, // no gate
++                    model.layers[il].ffn_down_exps,
++                    model.layers[il].ffn_exp_probs_b,
++                    n_expert, n_expert_used,
++                    LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
++                    true, hparams.expert_weights_scale,
++                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
++                    il);
++        cb(moe_out, "ffn_moe_out", il);
++
++        ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
++                    model.layers[il].ffn_up_shexp,  NULL, NULL,
++                    NULL /* no gate */           ,  NULL, NULL,
++                    model.layers[il].ffn_down_shexp, NULL, NULL,
++                    NULL,
++                    LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
++        cb(ffn_shexp, "ffn_shexp", il);
++
++        cur = ggml_add(ctx0, moe_out, ffn_shexp);
++        cb(cur, "ffn_out", il);
++    }
+ 
+     cur = build_cvec(cur, il);
+     cb(cur, "l_out", il);

llm/server.go

@@ -188,6 +188,11 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 	if len(projectors) > 0 && llamaModel != nil {
 		loadRequest.ProjectorPath = projectors[0]
 	}
+	// Determine if the user has forced FA on or off
+	faUserSet := false
+	if envconfig.FlashAttention(true) == envconfig.FlashAttention(false) {
+		faUserSet = true
+	}
 
 	fa := envconfig.FlashAttention(f.FlashAttention())
 
@@ -205,19 +210,51 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 
 	kvct := strings.ToLower(envconfig.KvCacheType())
 
-	if fa {
-		slog.Info("enabling flash attention")
-		loadRequest.FlashAttention = true
-
-		// Flash Attention also supports kv cache quantization
-		// Enable if the requested and kv cache type is supported by the model
-		if f.SupportsKVCacheType(kvct) {
-			loadRequest.KvCacheType = kvct
-		} else {
-			slog.Warn("kv cache type not supported by model", "type", kvct)
+	if textProcessor == nil {
+		flashAttention := ml.FlashAttentionAuto
+		if faUserSet {
+			if fa {
+				flashAttention = ml.FlashAttentionEnabled
+			} else {
+				flashAttention = ml.FlashAttentionDisabled
+			}
+		}
+
+		if kvct != "" {
+			if f.KVCacheTypeIsQuantized(kvct) {
+				if flashAttention != ml.FlashAttentionEnabled {
+					slog.Warn("OLLAMA_FLASH_ATTENTION must be enabled to use a quantized OLLAMA_KV_CACHE_TYPE", "type", kvct)
+					loadRequest.KvCacheType = ""
+				} else if f.SupportsKVCacheType(kvct) {
+					loadRequest.KvCacheType = kvct
+				} else {
+					slog.Warn("unsupported OLLAMA_KV_CACHE_TYPE", "type", kvct)
+				}
+			} else {
+				if f.SupportsKVCacheType(kvct) {
+					loadRequest.KvCacheType = kvct
+				} else {
+					slog.Warn("unsupported OLLAMA_KV_CACHE_TYPE", "type", kvct)
+				}
+			}
+		}
+		loadRequest.FlashAttention = flashAttention
+	} else {
+		// For Ollama engine, use our SupportsFlashAttention logic
+		if fa {
+			slog.Info("enabling flash attention")
+			loadRequest.FlashAttention = ml.FlashAttentionEnabled
+
+			// Flash Attention also supports kv cache quantization
+			// Enable if the requested and kv cache type is supported by the model
+			if f.SupportsKVCacheType(kvct) {
+				loadRequest.KvCacheType = kvct
+			} else {
+				slog.Warn("kv cache type not supported by model", "type", kvct)
+			}
+		} else if kvct != "" && kvct != "f16" {
+			slog.Warn("quantized kv cache requested but flash attention disabled", "type", kvct)
 		}
-	} else if kvct != "" && kvct != "f16" {
-		slog.Warn("quantized kv cache requested but flash attention disabled", "type", kvct)
 	}
 
 	gpuLibs := ml.LibraryPaths(gpus)
@@ -435,7 +472,7 @@ type LoadRequest struct {
 	LoraPath       []string
 	Parallel       int
 	BatchSize      int
-	FlashAttention bool
+	FlashAttention ml.FlashAttentionType
 	KvSize         int
 	KvCacheType    string
 	NumThreads     int
@@ -474,6 +511,13 @@ func (s *llamaServer) Load(ctx context.Context, systemInfo ml.SystemInfo, system
 		s.mem.GPUs[i].Cache = make([]uint64, s.totalLayers)
 	}
 
+	// Check if embedding model and adjust batch size accordingly
+	_, isEmbedding := s.ggml.KV()[fmt.Sprintf("%s.pooling_type", s.ggml.KV().Architecture())]
+	if isEmbedding && s.loadRequest.BatchSize < s.options.NumCtx {
+		s.loadRequest.BatchSize = s.options.NumCtx
+		slog.Info("embedding model detected, setting batch size to context length", "batch_size", s.loadRequest.BatchSize)
+	}
+
 	kv, graphPartialOffload, graphFullOffload := s.ggml.GraphSize(uint64(s.options.NumCtx), uint64(s.loadRequest.BatchSize),
 		s.loadRequest.Parallel, s.loadRequest.KvCacheType, s.loadRequest.FlashAttention)
 

middleware/openai.go

@@ -433,3 +433,111 @@ func ChatMiddleware() gin.HandlerFunc {
 		c.Next()
 	}
 }
+
+type ResponsesWriter struct {
+	BaseWriter
+	converter  *openai.ResponsesStreamConverter
+	model      string
+	stream     bool
+	responseID string
+	itemID     string
+}
+
+func (w *ResponsesWriter) writeEvent(eventType string, data any) error {
+	d, err := json.Marshal(data)
+	if err != nil {
+		return err
+	}
+	_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("event: %s\ndata: %s\n\n", eventType, d)))
+	if err != nil {
+		return err
+	}
+	if f, ok := w.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
+	}
+	return nil
+}
+
+func (w *ResponsesWriter) writeResponse(data []byte) (int, error) {
+	var chatResponse api.ChatResponse
+	if err := json.Unmarshal(data, &chatResponse); err != nil {
+		return 0, err
+	}
+
+	if w.stream {
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+
+		events := w.converter.Process(chatResponse)
+		for _, event := range events {
+			if err := w.writeEvent(event.Event, event.Data); err != nil {
+				return 0, err
+			}
+		}
+		return len(data), nil
+	}
+
+	// Non-streaming response
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	response := openai.ToResponse(w.model, w.responseID, w.itemID, chatResponse)
+	return len(data), json.NewEncoder(w.ResponseWriter).Encode(response)
+}
+
+func (w *ResponsesWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(data)
+	}
+	return w.writeResponse(data)
+}
+
+func ResponsesMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req openai.ResponsesRequest
+		if err := c.ShouldBindJSON(&req); err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		chatReq, err := openai.FromResponsesRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, openai.NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		// Check if client requested streaming (defaults to false)
+		streamRequested := req.Stream != nil && *req.Stream
+
+		// Pass streaming preference to the underlying chat request
+		chatReq.Stream = &streamRequested
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, openai.NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		responseID := fmt.Sprintf("resp_%d", rand.Intn(999999))
+		itemID := fmt.Sprintf("msg_%d", rand.Intn(999999))
+
+		w := &ResponsesWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			converter:  openai.NewResponsesStreamConverter(responseID, itemID, req.Model),
+			model:      req.Model,
+			stream:     streamRequested,
+			responseID: responseID,
+			itemID:     itemID,
+		}
+
+		// Set headers based on streaming mode
+		if streamRequested {
+			c.Writer.Header().Set("Content-Type", "text/event-stream")
+			c.Writer.Header().Set("Cache-Control", "no-cache")
+			c.Writer.Header().Set("Connection", "keep-alive")
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}

ml/backend.go

@@ -74,7 +74,7 @@ type BackendParams struct {
 	GPULayers GPULayersList
 
 	// FlashAttention indicates that we should use a fused flash attention kernel
-	FlashAttention bool
+	FlashAttention FlashAttentionType
 }
 
 var backends = make(map[string]func(string, BackendParams) (Backend, error))

ml/backend/ggml/ggml.go

@@ -109,7 +109,7 @@ type Backend struct {
 	// btDeviceMemory maps from a buffer type to the memory allocations associated with that device
 	btDeviceMemory map[C.ggml_backend_buffer_type_t]*ml.DeviceMemory
 
-	flashAttention bool
+	flashAttention ml.FlashAttentionType
 
 	// maxGraphNodes is the maximum allowed number of graph nodes in this scheduler
 	maxGraphNodes int
@@ -684,7 +684,7 @@ func (b *Backend) NewContextSize(n int) ml.Context {
 }
 
 func (b *Backend) CacheConfig() ml.CacheConfig {
-	if b.flashAttention {
+	if b.flashAttention == ml.FlashAttentionEnabled {
 		return ml.CacheConfig{CachePadding: 256, MaskDType: ml.DTypeF16, MaskBatchPadding: C.GGML_KQ_MASK_PAD}
 	} else {
 		return ml.CacheConfig{CachePadding: 256, PermutedV: true}
@@ -1534,7 +1534,8 @@ func (t *Tensor) RoPE(ctx ml.Context, positions ml.Tensor, ropeDim int, ropeBase
 			unsafe.SliceData(mropeSections),
 			C.int(opts.Type),
 			cmp.Or(C.int(opts.YaRN.OriginalContextLength), 128<<10),
-			C.float(ropeBase), C.float(ropeScale),
+			C.float(ropeBase),
+			C.float(ropeScale),
 			C.float(opts.YaRN.ExtrapolationFactor),
 			cmp.Or(C.float(opts.YaRN.AttentionFactor), 1),
 			cmp.Or(C.float(opts.YaRN.BetaFast), 32),
@@ -1546,9 +1547,11 @@ func (t *Tensor) RoPE(ctx ml.Context, positions ml.Tensor, ropeDim int, ropeBase
 			dequant,
 			positions.(*Tensor).t,
 			opts.Factors.(*Tensor).t,
-			C.int(ropeDim), C.int(opts.Type),
+			C.int(ropeDim),
+			C.int(opts.Type),
 			cmp.Or(C.int(opts.YaRN.OriginalContextLength), 128<<10),
-			C.float(ropeBase), C.float(ropeScale),
+			C.float(ropeBase),
+			C.float(ropeScale),
 			C.float(opts.YaRN.ExtrapolationFactor),
 			cmp.Or(C.float(opts.YaRN.AttentionFactor), 1),
 			cmp.Or(C.float(opts.YaRN.BetaFast), 32),
@@ -1676,7 +1679,7 @@ func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask, sin
 	query := t.Permute(ctx, 0, 2, 1, 3)
 	key = key.Permute(ctx, 0, 2, 1, 3)
 
-	if t.b.flashAttention {
+	if t.b.flashAttention == ml.FlashAttentionEnabled {
 		value = value.Permute(ctx, 0, 2, 1, 3)
 
 		kqv := C.ggml_flash_attn_ext(ctx.(*Context).ctx, query.(*Tensor).t, key.(*Tensor).t, value.(*Tensor).t, kqMask, C.float(scale), 0, 0)

ml/device.go

@@ -492,6 +492,32 @@ func FlashAttentionSupported(l []DeviceInfo) bool {
 	return true
 }
 
+type FlashAttentionType int32
+
+const (
+	// Aligned with llama_flash_attn_type
+	FlashAttentionAuto     FlashAttentionType = -1
+	FlashAttentionDisabled FlashAttentionType = 0
+	FlashAttentionEnabled  FlashAttentionType = 1
+)
+
+func (f FlashAttentionType) LogValue() slog.Value {
+	return slog.AnyValue(f.String())
+}
+
+func (f FlashAttentionType) String() string {
+	switch f {
+	case FlashAttentionAuto:
+		return "Auto"
+	case FlashAttentionDisabled:
+		return "Disabled"
+	case FlashAttentionEnabled:
+		return "Enabled"
+	default:
+		return "unknown"
+	}
+}
+
 // Given the list of GPUs this instantiation is targeted for,
 // figure out the visible devices environment variables
 // Set mustFilter true to enable filtering of CUDA devices

ml/nn/rope/options.go

@@ -77,6 +77,13 @@ func WithMRoPE(sections []int) func(*Options) {
 	}
 }
 
+func WithVision(sections []int) func(*Options) {
+	return func(opts *Options) {
+		opts.Type |= 1<<3 | 1<<4
+		opts.MRoPE.Sections = sections
+	}
+}
+
 func WithInterleaveMRoPE(sections []int) func(*Options) {
 	return func(opts *Options) {
 		opts.Type |= 1<<3 | 1<<5

model/models/gemma3/model_text.go

@@ -2,7 +2,6 @@ package gemma3
 
 import (
 	"math"
-	"slices"
 
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/kvcache"
@@ -13,25 +12,26 @@ import (
 )
 
 type TextConfig struct {
-	hiddenSize, numHeads, numKVHeads int
-	attnKeyLen, attnValLen           int
-	eps, ropeScale                   float32
-	ropeLocalBase                    float32
-	largeModelScaling                bool
-	slidingWindowPattern             []bool
-	ropeBase                         float32
-	ropeType                         string
-	ropeOriginalContext              int
-	ropeExtrapolation                float32
-	ropeBetaFast                     float32
-	ropeBetaSlow                     float32
-	finalLogitSoftcap                float32
+	hiddenSize, contextLength, numHeads, numKVHeads int
+	attnKeyLen, attnValLen                          int
+	eps, ropeScale                                  float32
+	ropeLocalBase                                   float32
+	largeModelScaling                               bool
+	slidingWindow                                   uint32
+	slidingWindowPattern                            []bool
+	ropeBase                                        float32
+	ropeType                                        string
+	ropeOriginalContext                             int
+	ropeExtrapolation                               float32
+	ropeBetaFast                                    float32
+	ropeBetaSlow                                    float32
+	finalLogitSoftcap                               float32
 }
 
-func (o TextConfig) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor, base float32) ml.Tensor {
+func (o TextConfig) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor, base, scale float32) ml.Tensor {
 	ropeOpts := []func(*rope.Options){rope.WithTypeNeoX()}
 	if o.ropeType == "yarn" {
-		attnFactor := float32(1.0 / (1.0 + 0.1*math.Log(float64(o.ropeScale))))
+		attnFactor := float32(1.0 / (1.0 + 0.1*math.Log(float64(scale))))
 		ropeOpts = append(ropeOpts,
 			rope.WithOriginalContextLength(o.ropeOriginalContext),
 			rope.WithExtrapolationFactor(o.ropeExtrapolation),
@@ -41,7 +41,7 @@ func (o TextConfig) applyRotaryPositionEmbeddings(ctx ml.Context, states, positi
 		)
 	}
 
-	return nn.RoPE(ctx, states, positions, o.attnKeyLen, base, 1./o.ropeScale, ropeOpts...)
+	return nn.RoPE(ctx, states, positions, o.attnKeyLen, base, 1./scale, ropeOpts...)
 }
 
 type TextModel struct {
@@ -55,6 +55,9 @@ type TextModel struct {
 
 const (
 	gemmaGlobalCacheCount = 6
+	gemma1BLayerCount     = 26
+	gemma4BLayerCount     = 34
+	gemma12BLayerCount    = 48
 	gemma27BLayerCount    = 62
 )
 
@@ -70,6 +73,7 @@ func newTextModel(c fs.Config) *TextModel {
 		Layers: make([]TextLayer, numBlocks),
 		TextConfig: &TextConfig{
 			hiddenSize:           int(c.Uint("embedding_length")),
+			contextLength:        int(c.Uint("context_length")),
 			numHeads:             int(c.Uint("attention.head_count")),
 			numKVHeads:           int(c.Uint("attention.head_count_kv")),
 			attnKeyLen:           int(c.Uint("attention.key_length", 256)),
@@ -77,6 +81,7 @@ func newTextModel(c fs.Config) *TextModel {
 			eps:                  c.Float("attention.layer_norm_rms_epsilon", 1e-06),
 			ropeLocalBase:        c.Float("rope.local.freq_base", 10000.0),
 			ropeBase:             c.Float("rope.freq_base", 1000000.0),
+			slidingWindow:        c.Uint("attention.sliding_window"),
 			slidingWindowPattern: c.Bools("attention.sliding_window_pattern"),
 			ropeType:             c.String("rope.scaling.type"),
 			ropeOriginalContext:  int(c.Uint("rope.scaling.original_context_length")),
@@ -88,14 +93,20 @@ func newTextModel(c fs.Config) *TextModel {
 		},
 	}
 
-	// Google's Gemma 3 release with sliding window attention does
-	// not use final logit softcapping, and so force it to 0.0
-	// TODO (jmorganca): this should ideally be set to 0.0 in the
-	// model configuration instead of here, as future versions of
-	// models may include both sliding window attention and final
-	// logit softcapping.
-	if slices.Contains(m.TextConfig.slidingWindowPattern, true) {
-		m.TextConfig.finalLogitSoftcap = 0.0
+	// Apply corrections for older versions of the Gemma 3 models
+	// by looking at whether they use sliding window attention and
+	// based on their layer counts.
+	if m.TextConfig.slidingWindow < uint32(m.TextConfig.contextLength) {
+		switch numBlocks {
+		case gemma1BLayerCount:
+			// The 1B model has final logit softcapping set to 30.0
+			// but it should be 0.0
+			m.TextConfig.finalLogitSoftcap = 0.0
+		case gemma4BLayerCount, gemma12BLayerCount, gemma27BLayerCount:
+			// The 4B, 12B, and 27B models have rope scale unset
+			// but it shuold be set to 8.0
+			m.TextConfig.ropeScale = 8.0
+		}
 	}
 
 	if numBlocks == gemma27BLayerCount {
@@ -114,31 +125,31 @@ type TextSelfAttention struct {
 	Output    *nn.Linear  `gguf:"attn_output"`
 }
 
-func (opts *TextConfig) ropeBaseForLayer(layer int) float32 {
+func (opts *TextConfig) ropeValuesForLayer(layer int) (base float32, scale float32) {
 	if opts.slidingWindowPattern != nil && opts.slidingWindowPattern[layer] {
-		return opts.ropeLocalBase
+		return opts.ropeLocalBase, 1.0
 	}
 
 	// Standard Gemma3: only every n-th layer is global,
 	// where n = gemmaGlobalCacheCount, otherwise use
 	// the local rope base
 	if (layer+1)%gemmaGlobalCacheCount > 0 {
-		return opts.ropeLocalBase
+		return opts.ropeLocalBase, 1.0
 	}
 
 	// default to global rope base
-	return opts.ropeBase
+	return opts.ropeBase, opts.ropeScale
 }
 
 func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *TextConfig) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 
-	ropeBase := opts.ropeBaseForLayer(layer)
+	ropeBase, ropeScale := opts.ropeValuesForLayer(layer)
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
 	q = sa.QueryNorm.Forward(ctx, q, opts.eps)
-	q = opts.applyRotaryPositionEmbeddings(ctx, q, positionIDs, ropeBase)
+	q = opts.applyRotaryPositionEmbeddings(ctx, q, positionIDs, ropeBase, ropeScale)
 
 	if opts.largeModelScaling {
 		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
@@ -149,7 +160,7 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
 	k = sa.KeyNorm.Forward(ctx, k, opts.eps)
-	k = opts.applyRotaryPositionEmbeddings(ctx, k, positionIDs, ropeBase)
+	k = opts.applyRotaryPositionEmbeddings(ctx, k, positionIDs, ropeBase, ropeScale)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
@@ -162,7 +173,8 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, pos
 }
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return m.applyRotaryPositionEmbeddings(ctx, key, shift, m.TextConfig.ropeBaseForLayer(layer)), nil
+	ropeBase, ropeScale := m.TextConfig.ropeValuesForLayer(layer)
+	return m.applyRotaryPositionEmbeddings(ctx, key, shift, ropeBase, ropeScale), nil
 }
 
 type TextMLP struct {

model/models/mistral3/model_text.go

@@ -8,6 +8,7 @@ import (
 	"github.com/ollama/ollama/kvcache"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model/input"
 )
 
@@ -17,10 +18,30 @@ type TextOptions struct {
 	eps, ropeBase, ropeScale         float32
 	ropeOrigPosEmbeddings            int
 	ropeScalingBeta                  float32
+	ropeType                         string
+	ropeExtrapolation                float32
+	ropeBetaFast                     float32
+	ropeBetaSlow                     float32
+	ropeMscale                       float32
+	ropeMscaleAllDim                 float32
 }
 
 func (o TextOptions) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
-	return nn.RoPE(ctx, states, positions, o.ropeDim, o.ropeBase, 1./o.ropeScale)
+	var ropeOpts []func(*rope.Options)
+	if o.ropeType == "yarn" {
+		if o.ropeMscale != 0 && o.ropeMscaleAllDim != 0 {
+			ropeOpts = append(ropeOpts, rope.WithAttentionFactor(1.0/float32(0.1*math.Log(float64(o.ropeScale))+1.0)))
+		}
+
+		ropeOpts = append(ropeOpts,
+			rope.WithOriginalContextLength(o.ropeOrigPosEmbeddings),
+			rope.WithExtrapolationFactor(o.ropeExtrapolation),
+			rope.WithBetaFast(o.ropeBetaFast),
+			rope.WithBetaSlow(o.ropeBetaSlow),
+		)
+	}
+
+	return nn.RoPE(ctx, states, positions, o.ropeDim, o.ropeBase, 1./o.ropeScale, ropeOpts...)
 }
 
 type TextModel struct {
@@ -150,9 +171,15 @@ func newTextModel(c fs.Config) *TextModel {
 			ropeDim:               int(c.Uint("rope.dimension_count")),
 			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:              c.Float("rope.freq_base"),
-			ropeScale:             c.Float("rope.scaling.factor", 1),
+			ropeScale:             c.Float("rope.scaling.factor", 1.0),
 			ropeOrigPosEmbeddings: int(c.Uint("rope.scaling.original_context_length")),
-			ropeScalingBeta:       c.Float("rope.scaling_beta"),
+			ropeScalingBeta:       c.Float("rope.scaling_beta", 0.1),
+			ropeBetaFast:          c.Float("rope.scaling.beta_fast", 32.0),
+			ropeBetaSlow:          c.Float("rope.scaling.beta_slow", 1.0),
+			ropeType:              c.String("rope.scaling.type"),
+			ropeMscale:            c.Float("rope.scaling.mscale"),
+			ropeMscaleAllDim:      c.Float("rope.scaling.mscale_all_dim"),
+			ropeExtrapolation:     c.Float("rope.scaling.extrapolation_factor", 1),
 		},
 	}
 }

model/models/models.go

@@ -13,6 +13,7 @@ import (
 	_ "github.com/ollama/ollama/model/models/mistral3"
 	_ "github.com/ollama/ollama/model/models/mllama"
 	_ "github.com/ollama/ollama/model/models/nomicbert"
+	_ "github.com/ollama/ollama/model/models/olmo3"
 	_ "github.com/ollama/ollama/model/models/qwen2"
 	_ "github.com/ollama/ollama/model/models/qwen25vl"
 	_ "github.com/ollama/ollama/model/models/qwen3"

model/models/olmo3/model.go

@@ -0,0 +1,223 @@
+package olmo3
+
+import (
+	"fmt"
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/rope"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+const (
+	cacheTypeSWA    = 0
+	cacheTypeCausal = 1
+)
+
+type Options struct {
+	hiddenSize, numHeads, numKVHeads int
+	eps, ropeBase, ropeScale         float32
+
+	originalContextLength int
+	attnFactor            float32
+
+	ropeType          string
+	ropeExtrapolation float32
+
+	slidingWindowPattern []bool
+}
+
+type Model struct {
+	model.Base
+	model.TextProcessor
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []Layer       `gguf:"blk"`
+	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
+	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
+
+	Options
+}
+
+func New(c fs.Config) (model.Model, error) {
+	vocabulary := model.Vocabulary{
+		Values: c.Strings("tokenizer.ggml.tokens"),
+		Scores: c.Floats("tokenizer.ggml.scores"),
+		Types:  c.Ints("tokenizer.ggml.token_type"),
+		Merges: c.Strings("tokenizer.ggml.merges"),
+		AddBOS: c.Bool("tokenizer.ggml.add_bos_token", false),
+		BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+		AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+		EOS: append(
+			[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+			c.Ints("tokenizer.ggml.eos_token_ids")...,
+		),
+	}
+
+	processor := model.NewBytePairEncoding(
+		&vocabulary,
+		"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+	)
+
+	m := Model{
+		TextProcessor: processor,
+		Layers:        make([]Layer, c.Uint("block_count")),
+		Options: Options{
+			hiddenSize:            int(c.Uint("embedding_length")),
+			numHeads:              int(c.Uint("attention.head_count")),
+			numKVHeads:            int(c.Uint("attention.head_count_kv")),
+			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:              c.Float("rope.freq_base", 1e4),
+			ropeScale:             c.Float("rope.scaling.factor", 1),
+			originalContextLength: int(c.Uint("rope.scaling.original_context_length")),
+			attnFactor:            c.Float("rope.scaling.attn_factor", 1),
+			ropeType:              c.String("rope.scaling.type"),
+			ropeExtrapolation:     c.Float("rope.scaling.extrapolation_factor", 1.0),
+			slidingWindowPattern:  c.Bools("attention.sliding_window_pattern"),
+		},
+	}
+
+	m.Cache = kvcache.NewWrapperCache(
+		kvcache.NewSWACache(int32(c.Uint("attention.sliding_window")), m.Shift),
+		kvcache.NewCausalCache(m.Shift),
+	)
+
+	return &m, nil
+}
+
+type SelfAttention struct {
+	Query  *nn.Linear  `gguf:"attn_q"`
+	Key    *nn.Linear  `gguf:"attn_k"`
+	Value  *nn.Linear  `gguf:"attn_v"`
+	Output *nn.Linear  `gguf:"attn_output"`
+	QNorm  *nn.RMSNorm `gguf:"attn_q_norm"`
+	KNorm  *nn.RMSNorm `gguf:"attn_k_norm"`
+}
+
+func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor, isSWA bool) ml.Tensor {
+	freqScale := float32(1.0)
+	ropeOpts := []func(*rope.Options){rope.WithTypeNeoX()}
+
+	if !isSWA {
+		freqScale = 1. / o.ropeScale
+		if o.originalContextLength > 0 {
+			ropeOpts = append(ropeOpts,
+				rope.WithOriginalContextLength(o.originalContextLength),
+				rope.WithExtrapolationFactor(o.ropeExtrapolation),
+			)
+		}
+	}
+
+	return nn.RoPE(ctx, states, positions, o.hiddenSize/o.numHeads, o.ropeBase, freqScale, ropeOpts...)
+}
+
+func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tensor, cache kvcache.Cache, m *Model, isSWA bool) ml.Tensor {
+	batchSize := hiddenState.Dim(1)
+	headDim := m.hiddenSize / m.numHeads
+
+	query := sa.Query.Forward(ctx, hiddenState)
+	query = sa.QNorm.Forward(ctx, query, m.eps)
+	query = query.Reshape(ctx, headDim, m.numHeads, batchSize)
+	query = m.Options.applyRotaryPositionEmbeddings(ctx, query, positions, isSWA)
+
+	key := sa.Key.Forward(ctx, hiddenState)
+	key = sa.KNorm.Forward(ctx, key, m.eps)
+	key = key.Reshape(ctx, headDim, m.numKVHeads, batchSize)
+	key = m.Options.applyRotaryPositionEmbeddings(ctx, key, positions, isSWA)
+
+	value := sa.Value.Forward(ctx, hiddenState)
+	value = value.Reshape(ctx, headDim, m.numKVHeads, batchSize)
+
+	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
+	attention = attention.Reshape(ctx, m.hiddenSize, batchSize)
+
+	return sa.Output.Forward(ctx, attention)
+}
+
+func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	isSWA := m.isSWALayer(layer)
+	return m.Options.applyRotaryPositionEmbeddings(ctx, key, shift, isSWA), nil
+}
+
+type MLP struct {
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+	Gate *nn.Linear `gguf:"ffn_gate"`
+}
+
+func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, m *Model) ml.Tensor {
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	return mlp.Down.Forward(ctx, hiddenState)
+}
+
+type Layer struct {
+	SelfAttention     *SelfAttention
+	PostAttentionNorm *nn.RMSNorm `gguf:"post_attention_norm"`
+	MLP               *MLP
+	PostFFWNorm       *nn.RMSNorm `gguf:"post_ffw_norm"`
+}
+
+func (l *Layer) Forward(ctx ml.Context, hiddenState, positions, outputs ml.Tensor, cache kvcache.Cache, m *Model, isSWA bool) ml.Tensor {
+	residual := hiddenState
+
+	hiddenState = l.SelfAttention.Forward(ctx, hiddenState, positions, cache, m, isSWA)
+
+	if outputs != nil {
+		hiddenState = hiddenState.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+	hiddenState = l.PostAttentionNorm.Forward(ctx, hiddenState, m.eps)
+
+	hiddenState = hiddenState.Add(ctx, residual)
+	residual = hiddenState
+
+	hiddenState = l.MLP.Forward(ctx, hiddenState, m)
+	hiddenState = l.PostFFWNorm.Forward(ctx, hiddenState, m.eps)
+
+	return hiddenState.Add(ctx, residual)
+}
+
+// OLMo3 has Sliding Window Attention (SWA) for 3 out of every 4 layers.
+func (m *Model) isSWALayer(layerIdx int) bool {
+	return m.Options.slidingWindowPattern[layerIdx]
+}
+
+func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
+
+	hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)
+
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+		cacheType := cacheTypeSWA
+
+		isSWA := m.isSWALayer(i)
+		if !isSWA {
+			cacheType = cacheTypeCausal
+		}
+
+		wc, ok := m.Cache.(*kvcache.WrapperCache)
+		if !ok {
+			return nil, fmt.Errorf("expected *kvcache.WrapperCache, got %T", m.Cache)
+		}
+		wc.SetLayerType(cacheType)
+
+		var outputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			outputs = batch.Outputs
+		}
+
+		hiddenState = layer.Forward(ctx, hiddenState, positions, outputs, m.Cache, m, isSWA)
+	}
+
+	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
+	return m.Output.Forward(ctx, hiddenState), nil
+}
+
+func init() {
+	model.Register("olmo3", New)
+}

model/models/qwen25vl/model.go

@@ -2,7 +2,6 @@ package qwen25vl
 
 import (
 	"bytes"
-	"fmt"
 	"image"
 	"slices"
 
@@ -33,7 +32,7 @@ func New(c fs.Config) (model.Model, error) {
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),
-				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", false),
 				BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
 				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
 				EOS: append(
@@ -54,19 +53,18 @@ func New(c fs.Config) (model.Model, error) {
 }
 
 func (m *Model) PixelValues(ctx ml.Context, multimodalData []byte) (ml.Tensor, *Grid, error) {
-	image, _, err := image.Decode(bytes.NewReader(multimodalData))
+	img, _, err := image.Decode(bytes.NewReader(multimodalData))
 	if err != nil {
 		return nil, nil, err
 	}
 
-	f32s, grid, err := m.ImageProcessor.ProcessImage(image)
+	f32s, grid, err := m.ImageProcessor.ProcessImage(img)
 	if err != nil {
 		return nil, nil, err
 	}
 
 	// Calculate tensor dimensions
-	patchDim := m.ImageProcessor.numChannels * m.ImageProcessor.temporalPatchSize *
-		m.ImageProcessor.patchSize * m.ImageProcessor.patchSize
+	patchDim := m.numChannels * m.temporalPatchSize * m.patchSize * m.patchSize
 	numPatches := grid.Temporal * grid.Height * grid.Width
 
 	pixelValues := ctx.Input().FromFloats(f32s, patchDim, numPatches)
@@ -85,11 +83,13 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
 	}
 
 	visionOutputs := m.VisionModel.Forward(ctx, pixels, grid)
-	return []input.Multimodal{{Tensor: visionOutputs}}, nil
+	return []input.Multimodal{{Tensor: visionOutputs, Data: grid}}, nil
 }
 
 // PostTokenize arranges Qwen-2.5-VL's inputs for the forward pass
 func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
+	// Reset position cache
+	m.positionCache = m.positionCache[:0]
 	var result []*input.Input
 
 	var (
@@ -98,40 +98,37 @@ func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
 		visionEndToken   int32 = 151653
 	)
 
-	nImg := 0
+	appendInput := func(i *input.Input, p int) int {
+		result = append(result, i)
+		m.positionCache = append(m.positionCache, int32(p))
+		return p + 1
+	}
+
+	var p int
 	for _, inp := range inputs {
 		if inp.Multimodal == nil {
 			// If not a multimodal input, add it to the result unchanged
-			result = append(result, inp)
+			p = appendInput(inp, p)
 		} else {
-			// Adding the 'Picture' prefix is a hack, at the time of writing there is no way to prefix
-			// the image tokens with a prompt, so we add a prefix here
-			nImg++
-			pre, err := m.Encode(fmt.Sprintf(" Picture %d: ", nImg), true)
-			if err != nil {
-				return nil, fmt.Errorf("failed to encode image prompt: %w", err)
-			}
-			for i := range pre {
-				result = append(result, &input.Input{Token: pre[i]})
-			}
-
-			patchesPerChunk := inp.Multimodal[0].Tensor.Dim(1)
-
 			// First add the vision start token
-			result = append(result, &input.Input{Token: visionStartToken})
+			p = appendInput(&input.Input{Token: visionStartToken}, p)
 
 			// Add the image token with the multimodal tensor data at the first position
-			result = append(result, &input.Input{
+			tokensPerGrid := inp.Multimodal[0].Tensor.Dim(1)
+			appendInput(&input.Input{
 				Token:          imageToken,
 				Multimodal:     inp.Multimodal,
 				MultimodalHash: inp.MultimodalHash,
-				SameBatch:      patchesPerChunk,
-			})
+				SameBatch:      tokensPerGrid,
+			}, p)
 
 			// Add the placeholder tokens for the remaining positions (tokensPerGrid-1)
-			result = append(result, slices.Repeat([]*input.Input{{Token: imageToken}}, patchesPerChunk-1)...)
+			for range tokensPerGrid - 1 {
+				appendInput(&input.Input{Token: imageToken}, p)
+			}
 
-			result = append(result, &input.Input{Token: visionEndToken})
+			grid := inp.Multimodal[0].Data.(*Grid)
+			p = appendInput(&input.Input{Token: visionEndToken}, p+max(grid.Width/m.spatialMergeSize, grid.Height/m.spatialMergeSize))
 		}
 	}
 
@@ -139,9 +136,58 @@ func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
 }
 
 func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
-	positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
+	// Initial token embedding
+	hiddenStates := m.TokenEmbedding.Forward(ctx, batch.Inputs).Duplicate(ctx)
 
-	return m.TextModel.Forward(ctx, batch.Inputs, positions, batch.Outputs, batch, m.Cache)
+	positionSlice := func() [][]int32 {
+		s := [][]int32{
+			make([]int32, len(batch.Positions)),
+			make([]int32, len(batch.Positions)),
+			make([]int32, len(batch.Positions)),
+			make([]int32, len(batch.Positions)),
+		}
+		for i, position := range batch.Positions {
+			if position < int32(len(m.positionCache)) {
+				position = m.positionCache[position]
+			} else if len(m.positionCache) > 0 {
+				position = position - int32(len(m.positionCache)) + m.positionCache[len(m.positionCache)-1] + 1
+			}
+
+			s[0][i] = position
+			s[1][i] = position
+			s[2][i] = position
+		}
+		return s
+	}()
+
+	for _, mi := range batch.Multimodal {
+		img := mi.Multimodal[0].Tensor
+		ctx.Forward(img.Copy(ctx, hiddenStates.View(ctx, mi.Index*hiddenStates.Stride(1), img.Dim(0)*img.Dim(1))))
+		if grid, ok := mi.Multimodal[0].Data.(*Grid); ok {
+			for i := range img.Dim(1) {
+				w := grid.Width / m.spatialMergeSize
+				positionSlice[1][mi.Index+i] += int32(i / w)
+				positionSlice[2][mi.Index+i] += int32(i % w)
+			}
+		}
+	}
+
+	positions := ctx.Input().FromInts(slices.Concat(positionSlice...), len(positionSlice[0])*len(positionSlice))
+
+	// Process through transformer layers
+	for i, layer := range m.TextModel.Layers {
+		m.Cache.SetLayer(i)
+
+		var lastLayerOutputs ml.Tensor
+		if i == len(m.TextModel.Layers)-1 {
+			lastLayerOutputs = batch.Outputs
+		}
+
+		hiddenStates = layer.Forward(ctx, hiddenStates, positions, lastLayerOutputs, m.Cache, m.TextOptions)
+	}
+
+	hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.TextModel.eps)
+	return m.Output.Forward(ctx, hiddenStates), nil
 }
 
 func init() {

model/models/qwen25vl/model_text.go

@@ -8,20 +8,17 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/ml/nn/rope"
-	"github.com/ollama/ollama/model/input"
 )
 
 type TextOptions struct {
 	hiddenSize, numHeads, numKVHeads int
 	ropeDim, originalContextLength   int
 	eps, ropeBase, ropeScale         float32
+	mropeSections                    []int
 }
 
 func (o TextOptions) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
-	return nn.RoPE(ctx, states, positions, o.ropeDim, o.ropeBase, 1./o.ropeScale,
-		rope.WithOriginalContextLength(o.originalContextLength),
-		rope.WithTypeNeoX(),
-	)
+	return nn.RoPE(ctx, states, positions, o.ropeDim, o.ropeBase, 1./o.ropeScale, rope.WithMRoPE(o.mropeSections))
 }
 
 type TextModel struct {
@@ -31,6 +28,7 @@ type TextModel struct {
 	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
 
 	*TextOptions
+	positionCache []int32
 }
 
 func NewTextModel(c fs.Config) *TextModel {
@@ -45,6 +43,14 @@ func NewTextModel(c fs.Config) *TextModel {
 			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
 			ropeBase:              c.Float("rope.freq_base"),
 			ropeScale:             c.Float("rope.scaling.factor", 1),
+			mropeSections: func() []int {
+				sections := c.Ints("rope.mrope_section")
+				s := make([]int, len(sections))
+				for i, section := range sections {
+					s[i] = int(section)
+				}
+				return s
+			}(),
 		},
 	}
 
@@ -84,6 +90,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 
 // Shift applies rotary position embeddings to the key tensor for causal attention caching
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	m.positionCache = nil
 	return m.applyRotaryPositionEmbeddings(ctx, key, shift), nil
 }
 
@@ -130,28 +137,3 @@ func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs, outputs ml.Ten
 	hiddenState = l.MLP.Forward(ctx, hiddenState, opts)
 	return hiddenState.Add(ctx, residual)
 }
-
-func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor, batch input.Batch, cache kvcache.Cache) (ml.Tensor, error) {
-	// Initial token embedding
-	hiddenStates := m.TokenEmbedding.Forward(ctx, inputs).Duplicate(ctx)
-
-	for _, mi := range batch.Multimodal {
-		img := mi.Multimodal[0].Tensor
-		ctx.Forward(img.Copy(ctx, hiddenStates.View(ctx, mi.Index*hiddenStates.Stride(1), img.Dim(0)*img.Dim(1))))
-	}
-
-	// Process through transformer layers
-	for i, layer := range m.Layers {
-		cache.SetLayer(i)
-
-		var lastLayerOutputs ml.Tensor
-		if i == len(m.Layers)-1 {
-			lastLayerOutputs = outputs
-		}
-
-		hiddenStates = layer.Forward(ctx, hiddenStates, positions, lastLayerOutputs, cache, m.TextOptions)
-	}
-
-	hiddenStates = m.OutputNorm.Forward(ctx, hiddenStates, m.eps)
-	return m.Output.Forward(ctx, hiddenStates), nil
-}

model/models/qwen25vl/model_vision.go

@@ -7,48 +7,28 @@ import (
 	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/rope"
 )
 
-// We only support batch size of 1
-var batchSize int = 1
-
-func rotateHalf(ctx ml.Context, t ml.Tensor) ml.Tensor {
-	x1 := t.Slice(ctx, 0, 0, t.Dim(0)/2, 1)
-	x2 := t.Slice(ctx, 0, t.Dim(0)/2, t.Dim(0), 1).Contiguous(ctx)
-	return x2.Scale(ctx, -1).Concat(ctx, x1, 0)
-}
-
-func applyRotaryPositionEmbeddings(ctx ml.Context, states, cos, sin ml.Tensor) ml.Tensor {
-	return states.Mul(ctx, cos).Add(ctx, rotateHalf(ctx, states).Mul(ctx, sin))
-}
-
-func blockDiagonalMask(ctx ml.Context, seqLength int, bounds []int, numHeads int) ml.Tensor {
-	// Create a flat slice for the mask (all -inf initially to block all attention)
-	flat := make([]float32, seqLength*seqLength)
-	for i := range flat {
-		flat[i] = float32(math.Inf(-1)) // Negative infinity to block attention
+func blockDiagonalMask(ctx ml.Context, seqLength int, bounds []int) ml.Tensor {
+	// Initialize a 2D mask with -Inf
+	s := make([][]float32, seqLength)
+	for i := range s {
+		s[i] = slices.Repeat([]float32{float32(math.Inf(-1))}, seqLength)
 	}
 
 	// Fill in the mask with zeros for tokens that CAN attend to each other
 	for i := 1; i < len(bounds); i++ {
-		start := bounds[i-1]
-		end := bounds[i]
-
-		// Enable attention within this sequence block by setting values to 0
+		start, end := bounds[i-1], bounds[i]
+		// Enable attention within this sequence block
 		for row := start; row < end; row++ {
 			for col := start; col < end; col++ {
-				idx := row*seqLength + col
-				flat[idx] = 0.0 // 0 allows attention, -inf blocks it
+				s[row][col] = 0.0
 			}
 		}
 	}
 
-	mask := ctx.Input().FromFloats(flat, seqLength, seqLength)
-
-	// Reshape to match [seqLength, seqLength, 1] for broadcasting
-	mask = mask.Reshape(ctx, seqLength, seqLength, 1)
-
-	return mask
+	return ctx.Input().FromFloats(slices.Concat(s...), seqLength, seqLength)
 }
 
 type VisionSelfAttention struct {
@@ -58,17 +38,17 @@ type VisionSelfAttention struct {
 	Output *nn.Linear `gguf:"attn_out"`
 }
 
-func (sa *VisionSelfAttention) Forward(ctx ml.Context, hiddenStates, cos, sin, mask ml.Tensor, opts *VisionModelOptions) ml.Tensor {
+func (sa *VisionSelfAttention) Forward(ctx ml.Context, hiddenStates, positions, mask ml.Tensor, opts *VisionModelOptions) ml.Tensor {
 	query := sa.Query.Forward(ctx, hiddenStates)
 	key := sa.Key.Forward(ctx, hiddenStates)
 	value := sa.Value.Forward(ctx, hiddenStates)
 
-	query = query.Reshape(ctx, opts.headDim, opts.numHeads, query.Dim(1), batchSize)
-	key = key.Reshape(ctx, opts.headDim, opts.numHeads, key.Dim(1), batchSize)
-	value = value.Reshape(ctx, opts.headDim, opts.numHeads, value.Dim(1), batchSize)
+	query = query.Reshape(ctx, opts.headDim, opts.numHeads, query.Dim(1))
+	key = key.Reshape(ctx, opts.headDim, opts.numHeads, key.Dim(1))
+	value = value.Reshape(ctx, opts.headDim, opts.numHeads, value.Dim(1))
 
-	query = applyRotaryPositionEmbeddings(ctx, query, cos, sin)
-	key = applyRotaryPositionEmbeddings(ctx, key, cos, sin)
+	query = opts.applyRotaryPositionEmbeddings(ctx, query, positions)
+	key = opts.applyRotaryPositionEmbeddings(ctx, key, positions)
 
 	// Scale factor for scaled dot-product attention
 	scale := 1.0 / math.Sqrt(float64(opts.headDim))
@@ -77,6 +57,7 @@ func (sa *VisionSelfAttention) Forward(ctx ml.Context, hiddenStates, cos, sin, m
 	query = query.Permute(ctx, 0, 2, 1, 3)
 	key = key.Permute(ctx, 0, 2, 1, 3)
 	value = value.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx)
+
 	kq := key.MulmatFullPrec(ctx, query)
 	kq = kq.Scale(ctx, scale)
 	if mask != nil {
@@ -85,7 +66,7 @@ func (sa *VisionSelfAttention) Forward(ctx ml.Context, hiddenStates, cos, sin, m
 	kq = kq.Softmax(ctx)
 	kqv := value.Mulmat(ctx, kq)
 	attention := kqv.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
-	attention = attention.Reshape(ctx, opts.hiddenSize, attention.Dim(2), batchSize)
+	attention = attention.Reshape(ctx, opts.hiddenSize, attention.Dim(2))
 
 	return sa.Output.Forward(ctx, attention)
 }
@@ -98,10 +79,7 @@ type VisionMLP struct {
 }
 
 func (mlp *VisionMLP) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *VisionModelOptions) ml.Tensor {
-	// Using activation as specified in config (likely GELU or SiLU/Swish)
-	gateOutput := mlp.Gate.Forward(ctx, hiddenStates)
-	hiddenStates = gateOutput.SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
-
+	hiddenStates = mlp.Gate.Forward(ctx, hiddenStates).SILU(ctx, mlp.Up.Forward(ctx, hiddenStates))
 	return mlp.Down.Forward(ctx, hiddenStates)
 }
 
@@ -112,10 +90,10 @@ type VisionEncoderLayer struct {
 	MLP           *VisionMLP
 }
 
-func (e *VisionEncoderLayer) Forward(ctx ml.Context, hiddenStates, cos, sin, mask ml.Tensor, opts *VisionModelOptions) ml.Tensor {
+func (e *VisionEncoderLayer) Forward(ctx ml.Context, hiddenStates, positions, mask ml.Tensor, opts *VisionModelOptions) ml.Tensor {
 	residual := hiddenStates
 	hiddenStates = e.Norm1.Forward(ctx, hiddenStates, opts.eps)
-	hiddenStates = e.SelfAttention.Forward(ctx, hiddenStates, cos, sin, mask, opts)
+	hiddenStates = e.SelfAttention.Forward(ctx, hiddenStates, positions, mask, opts)
 	hiddenStates = hiddenStates.Add(ctx, residual)
 
 	residual = hiddenStates
@@ -139,6 +117,17 @@ type VisionModelOptions struct {
 	temporalPatchSize int
 }
 
+func (o VisionModelOptions) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions ml.Tensor) ml.Tensor {
+	return nn.RoPE(ctx, states, positions, o.headDim/2, o.ropeTheta, 1,
+		rope.WithVision([]int{
+			o.headDim / 4,
+			o.headDim / 4,
+			o.headDim / 4,
+			o.headDim / 4,
+		}),
+	)
+}
+
 type PatchEmbedding struct {
 	PatchConv0 *nn.Conv2D `gguf:"patch_embd_0"`
 	PatchConv1 *nn.Conv2D `gguf:"patch_embd_1"`
@@ -186,7 +175,7 @@ func (pm *VisionPatchMerger) Forward(ctx ml.Context, visionOutputs ml.Tensor, op
 	hiddenSize := visionOutputs.Dim(0) * (opts.spatialMergeSize * opts.spatialMergeSize)
 
 	// Reshape the normalized output to view the hidden size dimension
-	reshaped := normalized.Reshape(ctx, hiddenSize, normalized.Dim(1)/(opts.spatialMergeSize*opts.spatialMergeSize), batchSize)
+	reshaped := normalized.Reshape(ctx, hiddenSize, normalized.Dim(1)/(opts.spatialMergeSize*opts.spatialMergeSize))
 	hidden := pm.MLP0.Forward(ctx, reshaped)
 	activated := hidden.GELU(ctx)
 
@@ -209,36 +198,53 @@ func (m *VisionModel) Forward(ctx ml.Context, pixelValues ml.Tensor, grid *Grid)
 	// Extract patch embeddings
 	hiddenStates := m.PatchEmbedding.Forward(ctx, pixelValues, m.VisionModelOptions)
 
-	positionEmbedding := m.PositionalEmbedding(ctx, grid)
-
-	windowIndex, bounds := m.WindowIndex(ctx, grid)
-
+	index, bounds := m.windowIndex(grid)
 	spatialMergeUnit := m.spatialMergeSize * m.spatialMergeSize
 
+	windowIndex := ctx.Input().FromInts(index, len(index))
 	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0)*spatialMergeUnit, hiddenStates.Dim(1)/spatialMergeUnit)
-	hiddenStates = hiddenStates.Rows(ctx, windowIndex)
+	hiddenStates = hiddenStates.Rows(ctx, windowIndex.Argsort(ctx))
 	hiddenStates = hiddenStates.Reshape(ctx, hiddenStates.Dim(0)/spatialMergeUnit, hiddenStates.Dim(1)*spatialMergeUnit)
 
-	positionEmbedding = positionEmbedding.Reshape(ctx, positionEmbedding.Dim(0)*spatialMergeUnit, positionEmbedding.Dim(1)/spatialMergeUnit)
-	positionEmbedding = positionEmbedding.Rows(ctx, windowIndex)
-	positionEmbedding = positionEmbedding.Reshape(ctx, positionEmbedding.Dim(0)/spatialMergeUnit, positionEmbedding.Dim(1)*spatialMergeUnit)
-	positionEmbedding = positionEmbedding.Concat(ctx, positionEmbedding, 0)
+	positions := ctx.Input().FromInts(func() []int32 {
+		s := [][]int32{
+			make([]int32, grid.Height*grid.Width),
+			make([]int32, grid.Height*grid.Width),
+			make([]int32, grid.Height*grid.Width),
+			make([]int32, grid.Height*grid.Width),
+		}
 
-	cos, sin := positionEmbedding.Cos(ctx), positionEmbedding.Sin(ctx)
-	cos = cos.Reshape(ctx, cos.Dim(0), 1, cos.Dim(1))
-	sin = sin.Reshape(ctx, sin.Dim(0), 1, sin.Dim(1))
+		var cur int
+		for y := 0; y < grid.Height; y += m.spatialMergeSize {
+			for x := 0; x < grid.Width; x += m.spatialMergeSize {
+				for dy := range 2 {
+					for dx := range 2 {
+						i := int(index[cur/spatialMergeUnit]) * spatialMergeUnit
+						i += cur % spatialMergeUnit
+						s[0][i] = int32(y + dy)
+						s[1][i] = int32(x + dx)
+						s[2][i] = int32(y + dy)
+						s[3][i] = int32(x + dx)
+						cur++
+					}
+				}
+			}
+		}
+
+		return slices.Concat(s...)
+	}(), grid.Height*grid.Width*4)
+
+	mask := blockDiagonalMask(ctx, hiddenStates.Dim(1), bounds)
 
-	mask := blockDiagonalMask(ctx, hiddenStates.Dim(1), bounds, m.VisionModelOptions.numHeads)
 	// Apply encoder layers
 	for i, layer := range m.Layers {
 		if slices.Contains(m.fullAttnBlocks, int32(i)) {
-			hiddenStates = layer.Forward(ctx, hiddenStates, cos, sin, nil, m.VisionModelOptions)
+			hiddenStates = layer.Forward(ctx, hiddenStates, positions, nil, m.VisionModelOptions)
 		} else {
 			hiddenStates = layer.Forward(
 				ctx,
 				hiddenStates,
-				cos,
-				sin,
+				positions,
 				mask,
 				m.VisionModelOptions,
 			)
@@ -246,102 +252,43 @@ func (m *VisionModel) Forward(ctx ml.Context, pixelValues ml.Tensor, grid *Grid)
 	}
 
 	hiddenStates = m.PatchMerger.Forward(ctx, hiddenStates, m.VisionModelOptions)
-	reverseWindowIndex := windowIndex.Argsort(ctx)
-	return hiddenStates.Rows(ctx, reverseWindowIndex)
+	return hiddenStates.Rows(ctx, windowIndex)
 }
 
-// WindowIndex divides the grid into windows and returns:
-//  1. A tensor containing flattened indices of all grid points organized by windows
+// windowIndex divides the grid into windows and returns:
+//  1. A slice of grid point indices organized by windows
 //  2. A slice of boundaries that mark where each window's data begins and ends
 //     in the flattened representation, scaled by spatialMergeSize squared
 //
 // The boundaries slice always starts with 0 and contains cumulative ending
 // positions for each window, allowing downstream processing to identify
 // window boundaries in the tensor data.
-func (m *VisionModel) WindowIndex(ctx ml.Context, grid *Grid) (ml.Tensor, []int) {
-	vitMergerWindowSize := m.windowSize / m.spatialMergeSize / m.patchSize
+func (m *VisionModel) windowIndex(grid *Grid) (index []int32, bounds []int) {
+	height := grid.Height / m.spatialMergeSize
+	width := grid.Width / m.spatialMergeSize
+	window := m.windowSize / m.patchSize / m.spatialMergeSize
 
-	llmGridH := grid.Height / m.spatialMergeSize
-	llmGridW := grid.Width / m.spatialMergeSize
+	index = make([]int32, height*width)
 
-	// Calculate window parameters
-	numWindowsH := int(math.Ceil(float64(llmGridH) / float64(vitMergerWindowSize)))
-	numWindowsW := int(math.Ceil(float64(llmGridW) / float64(vitMergerWindowSize)))
+	bounds = make([]int, 0, ((height+window-1)/window)*((width+window-1)/window)+1)
+	bounds = append(bounds, 0)
 
-	// Initialize index_new slice
-	var index []int32
-
-	// Initialize bounds with the first element as 0
-	bounds := []int{0}
-	totalSeqLen := 0
-
-	// Process each window without padding
-	for wh := range numWindowsH {
-		for ww := range numWindowsW {
-			// Calculate window boundaries
-			hStart := wh * vitMergerWindowSize
-			wStart := ww * vitMergerWindowSize
-			hEnd := min(hStart+vitMergerWindowSize, llmGridH)
-			wEnd := min(wStart+vitMergerWindowSize, llmGridW)
-
-			// Calculate sequence length for this window
-			seqLen := (hEnd - hStart) * (wEnd - wStart)
-
-			// Collect indices for this window
-			for h := hStart; h < hEnd; h++ {
-				for w := wStart; w < wEnd; w++ {
-					index = append(index, int32(h*llmGridW+w))
+	var cur int32
+	for y := 0; y < height; y += window {
+		for x := 0; x < width; x += window {
+			h1 := min(window, height-y)
+			w1 := min(window, width-x)
+			for dy := range h1 {
+				for dx := range w1 {
+					win := (y+dy)*width + (x + dx)
+					index[win] = cur
+					cur++
 				}
 			}
-
-			totalSeqLen += seqLen
-			bounds = append(bounds, totalSeqLen*(m.spatialMergeSize*m.spatialMergeSize)+bounds[0])
+			bounds = append(bounds, int(cur)*window)
 		}
 	}
-
-	t := ctx.Input().FromInts(index, len(index))
-
-	return t, bounds
-}
-
-// PositionalEmbedding generates rotary position embeddings for attention mechanisms
-func (m *VisionModel) PositionalEmbedding(ctx ml.Context, grid *Grid) ml.Tensor {
-	dim := m.headDim / 2
-	freq := dim / 2
-	theta := float64(m.ropeTheta)
-	merge := m.spatialMergeSize
-
-	// Create frequency patterns for position encoding
-	maxGridSize := max(grid.Height, grid.Width)
-	freqVals := make([]float32, freq*maxGridSize)
-	for i := range maxGridSize {
-		for j := range freq {
-			freqVals[i*freq+j] = float32(i) / float32(math.Pow(theta, float64(j*2)/float64(dim)))
-		}
-	}
-	freqs := ctx.Input().FromFloats(freqVals, freq, maxGridSize)
-
-	// Create position coordinates (y,x pairs) for the grid
-	// In PyTorch: Equivalent to generating position ids with torch.arange()
-	coords := make([]int32, 0, grid.Height*grid.Width*2)
-	for y := range grid.Height {
-		for x := range grid.Width {
-			coords = append(coords, int32(y), int32(x))
-		}
-	}
-	pos := ctx.Input().FromInts(coords, 2, grid.Width, grid.Height)
-
-	// Reshape and permute positions to match spatial merging pattern
-	pos = pos.Reshape(ctx, 2, grid.Width, merge, grid.Height/merge)
-	pos = pos.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
-	pos = pos.Reshape(ctx, 2, merge, merge, grid.Width/merge*grid.Height/merge)
-	pos = pos.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
-	pos = pos.Reshape(ctx, 2*merge*merge*grid.Width/merge*grid.Height/merge)
-
-	// Use position indices to look up corresponding frequency values
-	positionalEmbedding := freqs.Rows(ctx, pos)
-	positionalEmbedding = positionalEmbedding.Reshape(ctx, positionalEmbedding.Dim(0)*2, positionalEmbedding.Dim(1)/2)
-	return positionalEmbedding
+	return index, bounds
 }
 
 // newVisionModel creates a new instance of the Qwen vision model

model/models/qwen25vl/process_image.go

@@ -19,8 +19,8 @@ type ImageProcessor struct {
 	maxPixels         int
 	factor            int
 	rescaleFactor     float32
-	imageMean         []float32
-	imageStd          []float32
+	imageMean         [3]float32
+	imageStd          [3]float32
 }
 
 // newImageProcessor creates a new image processor with default values
@@ -34,11 +34,11 @@ func newImageProcessor(c fs.Config) ImageProcessor {
 		temporalPatchSize: 2,
 		mergeSize:         mergeSize,
 		minPixels:         56 * 56,
-		maxPixels:         int(c.Uint("vision.max_pixels", 28*28*1280)), // 1MP limit
+		maxPixels:         int(c.Uint("vision.max_pixels", 2<<20)), // 2M limit
 		factor:            patchSize * mergeSize,
 		rescaleFactor:     1.0 / 255.0,
-		imageMean:         imageproc.ClipDefaultMean[:],
-		imageStd:          imageproc.ClipDefaultSTD[:],
+		imageMean:         imageproc.ClipDefaultMean,
+		imageStd:          imageproc.ClipDefaultSTD,
 	}
 }
 
@@ -90,13 +90,7 @@ func (p *ImageProcessor) ProcessImage(img image.Image) ([]float32, *Grid, error)
 	// Resize image using existing functions
 	resizedImg := imageproc.Resize(img, image.Point{X: resizedWidth, Y: resizedHeight}, imageproc.ResizeBilinear)
 
-	normalizedPixels := imageproc.Normalize(
-		resizedImg,
-		[3]float32{p.imageMean[0], p.imageMean[1], p.imageMean[2]},
-		[3]float32{p.imageStd[0], p.imageStd[1], p.imageStd[2]},
-		true, // rescale
-		true, // channelFirst
-	)
+	normalizedPixels := imageproc.Normalize(resizedImg, p.imageMean, p.imageStd, true, true)
 
 	// Calculate grid dimensions
 	grid := &Grid{

model/parsers/nemotron3nano.go

@@ -0,0 +1,254 @@
+package parsers
+
+import (
+	"regexp"
+	"strings"
+	"unicode"
+
+	"github.com/ollama/ollama/api"
+)
+
+type Nemotron3NanoParserState int
+
+const (
+	Nemotron3NanoCollectingThinking Nemotron3NanoParserState = iota
+	Nemotron3NanoSkipWhitespaceAfterThinking
+	Nemotron3NanoCollectingContent
+	Nemotron3NanoCollectingToolCalls
+)
+
+const (
+	nemotronThinkClose    = "</think>"
+	nemotronToolCallOpen  = "<tool_call>"
+	nemotronToolCallClose = "</tool_call>"
+)
+
+type Nemotron3NanoParser struct {
+	state  Nemotron3NanoParserState
+	buffer strings.Builder
+	tools  []api.Tool
+}
+
+func (p *Nemotron3NanoParser) HasToolSupport() bool     { return true }
+func (p *Nemotron3NanoParser) HasThinkingSupport() bool { return true }
+
+func (p *Nemotron3NanoParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.tools = tools
+
+	// thinking is enabled if user requests it
+	thinkingEnabled := thinkValue != nil && thinkValue.Bool()
+
+	prefill := lastMessage != nil && lastMessage.Role == "assistant"
+
+	if !thinkingEnabled {
+		p.state = Nemotron3NanoCollectingContent
+		return tools
+	}
+
+	if prefill && lastMessage.Content != "" {
+		p.state = Nemotron3NanoCollectingContent
+		return tools
+	}
+
+	p.state = Nemotron3NanoCollectingThinking
+	return tools
+}
+
+type nemotronEvent interface {
+	isNemotronEvent()
+}
+
+type nemotronEventThinkingContent struct {
+	content string
+}
+
+type nemotronEventContent struct {
+	content string
+}
+
+type nemotronEventToolCall struct {
+	toolCall api.ToolCall
+}
+
+func (nemotronEventThinkingContent) isNemotronEvent() {}
+func (nemotronEventContent) isNemotronEvent()         {}
+func (nemotronEventToolCall) isNemotronEvent()        {}
+
+func (p *Nemotron3NanoParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	p.buffer.WriteString(s)
+	events := p.parseEvents()
+
+	var toolCalls []api.ToolCall
+	var contentSb strings.Builder
+	var thinkingSb strings.Builder
+	for _, event := range events {
+		switch event := event.(type) {
+		case nemotronEventToolCall:
+			toolCalls = append(toolCalls, event.toolCall)
+		case nemotronEventThinkingContent:
+			thinkingSb.WriteString(event.content)
+		case nemotronEventContent:
+			contentSb.WriteString(event.content)
+		}
+	}
+
+	return contentSb.String(), thinkingSb.String(), toolCalls, nil
+}
+
+func (p *Nemotron3NanoParser) parseEvents() []nemotronEvent {
+	var all []nemotronEvent
+
+	keepLooping := true
+	for keepLooping {
+		var events []nemotronEvent
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	return all
+}
+
+// emitWithPartialCheck extracts unambiguous content before a potential partial tag
+func (p *Nemotron3NanoParser) emitWithPartialCheck(bufStr, tag string) (unambiguous, ambiguous string) {
+	if overlapLen := overlap(bufStr, tag); overlapLen > 0 {
+		beforePartialTag := bufStr[:len(bufStr)-overlapLen]
+		trailingLen := trailingWhitespaceLen(beforePartialTag)
+		return bufStr[:len(beforePartialTag)-trailingLen], bufStr[len(beforePartialTag)-trailingLen:]
+	}
+	wsLen := trailingWhitespaceLen(bufStr)
+	return bufStr[:len(bufStr)-wsLen], bufStr[len(bufStr)-wsLen:]
+}
+
+func (p *Nemotron3NanoParser) eat() ([]nemotronEvent, bool) {
+	bufStr := p.buffer.String()
+	if bufStr == "" {
+		return nil, false
+	}
+
+	switch p.state {
+	case Nemotron3NanoCollectingThinking:
+		if strings.Contains(bufStr, nemotronThinkClose) {
+			split := strings.SplitN(bufStr, nemotronThinkClose, 2)
+			thinking := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			p.buffer.Reset()
+			remainder := strings.TrimLeftFunc(split[1], unicode.IsSpace)
+			p.buffer.WriteString(remainder)
+			// Transition to whitespace-skipping state if buffer is empty,
+			// otherwise go directly to content collection
+			if remainder == "" {
+				p.state = Nemotron3NanoSkipWhitespaceAfterThinking
+			} else {
+				p.state = Nemotron3NanoCollectingContent
+			}
+			if thinking != "" {
+				return []nemotronEvent{nemotronEventThinkingContent{content: thinking}}, true
+			}
+			return nil, true
+		}
+		unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronThinkClose)
+		p.buffer.Reset()
+		p.buffer.WriteString(ambig)
+		if unambig != "" {
+			return []nemotronEvent{nemotronEventThinkingContent{content: unambig}}, false
+		}
+		return nil, false
+
+	// We only want to skip whitespace between thinking and content
+	case Nemotron3NanoSkipWhitespaceAfterThinking:
+		bufStr = strings.TrimLeftFunc(bufStr, unicode.IsSpace)
+		p.buffer.Reset()
+		p.buffer.WriteString(bufStr)
+		if bufStr == "" {
+			return nil, false
+		}
+		p.state = Nemotron3NanoCollectingContent
+		return nil, true
+
+	case Nemotron3NanoCollectingContent:
+		if strings.Contains(bufStr, nemotronToolCallOpen) {
+			split := strings.SplitN(bufStr, nemotronToolCallOpen, 2)
+			content := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			p.buffer.Reset()
+			p.buffer.WriteString(split[1])
+			p.state = Nemotron3NanoCollectingToolCalls
+			if content != "" {
+				return []nemotronEvent{nemotronEventContent{content: content}}, true
+			}
+			return nil, true
+		}
+		unambig, ambig := p.emitWithPartialCheck(bufStr, nemotronToolCallOpen)
+		p.buffer.Reset()
+		p.buffer.WriteString(ambig)
+		if unambig != "" {
+			return []nemotronEvent{nemotronEventContent{content: unambig}}, false
+		}
+		return nil, false
+
+	case Nemotron3NanoCollectingToolCalls:
+		if strings.Contains(bufStr, nemotronToolCallClose) {
+			split := strings.SplitN(bufStr, nemotronToolCallClose, 2)
+			remaining := strings.TrimLeftFunc(split[1], unicode.IsSpace)
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+
+			var events []nemotronEvent
+			if tc, err := p.parseToolCall(split[0]); err == nil {
+				events = append(events, nemotronEventToolCall{toolCall: tc})
+			}
+
+			if !strings.Contains(remaining, nemotronToolCallOpen) {
+				p.state = Nemotron3NanoCollectingContent
+			}
+			return events, true
+		}
+		return nil, false
+	}
+
+	return nil, false
+}
+
+var (
+	nemotronFunctionRegex  = regexp.MustCompile(`<function=([^>]+)>`)
+	nemotronParameterRegex = regexp.MustCompile(`<parameter=([^>]+)>\n?([\s\S]*?)\n?</parameter>`)
+)
+
+func (p *Nemotron3NanoParser) parseToolCall(content string) (api.ToolCall, error) {
+	toolCall := api.ToolCall{}
+
+	// Extract function name
+	fnMatch := nemotronFunctionRegex.FindStringSubmatch(content)
+	if len(fnMatch) < 2 {
+		return toolCall, nil
+	}
+	toolCall.Function.Name = fnMatch[1]
+
+	// Extract parameters
+	toolCall.Function.Arguments = make(api.ToolCallFunctionArguments)
+	paramMatches := nemotronParameterRegex.FindAllStringSubmatch(content, -1)
+	for _, match := range paramMatches {
+		if len(match) >= 3 {
+			paramName := match[1]
+			paramValue := strings.TrimSpace(match[2])
+
+			// Try to parse as typed value based on tool definition
+			toolCall.Function.Arguments[paramName] = p.parseParamValue(paramName, paramValue)
+		}
+	}
+
+	return toolCall, nil
+}
+
+func (p *Nemotron3NanoParser) parseParamValue(paramName string, raw string) any {
+	// Find the matching tool to get parameter type
+	var paramType api.PropertyType
+	for _, tool := range p.tools {
+		if prop, ok := tool.Function.Parameters.Properties[paramName]; ok {
+			paramType = prop.Type
+			break
+		}
+	}
+
+	return parseValue(raw, paramType)
+}

model/parsers/nemotron3nano_test.go

@@ -0,0 +1,574 @@
+package parsers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestNemotron3NanoParser(t *testing.T) {
+	tests := []struct {
+		name             string
+		input            string
+		thinkValue       *api.ThinkValue
+		expectedContent  string
+		expectedThinking string
+		expectedCalls    []api.ToolCall
+	}{
+		{
+			name:            "simple content - no thinking",
+			input:           "Hello, how can I help you?",
+			thinkValue:      nil,
+			expectedContent: "Hello, how can I help you?",
+		},
+		{
+			name:            "simple content - thinking disabled",
+			input:           "Hello, how can I help you?",
+			thinkValue:      &api.ThinkValue{Value: false},
+			expectedContent: "Hello, how can I help you?",
+		},
+		{
+			name:             "thinking then content",
+			input:            "Let me think about this...</think>\nHere is my answer.",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Let me think about this...",
+			expectedContent:  "Here is my answer.",
+		},
+		{
+			name:             "thinking with newlines",
+			input:            "Step 1: Analyze\nStep 2: Process\nStep 3: Conclude</think>\nThe answer is 42.",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Step 1: Analyze\nStep 2: Process\nStep 3: Conclude",
+			expectedContent:  "The answer is 42.",
+		},
+		{
+			name:       "simple tool call",
+			input:      "<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "Paris"},
+					},
+				},
+			},
+		},
+		{
+			name:            "content then tool call",
+			input:           "Let me check the weather.\n<tool_call>\n<function=get_weather>\n<parameter=city>\nNYC\n</parameter>\n</function>\n</tool_call>",
+			thinkValue:      nil,
+			expectedContent: "Let me check the weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "NYC"},
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with multiple parameters",
+			input:      "<tool_call>\n<function=book_flight>\n<parameter=from>\nSFO\n</parameter>\n<parameter=to>\nNYC\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "book_flight",
+						Arguments: map[string]any{
+							"from": "SFO",
+							"to":   "NYC",
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "multiple tool calls",
+			input: "<tool_call>\n<function=get_weather>\n<parameter=city>\nSan Francisco\n</parameter>\n</function>\n</tool_call>\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nNew York\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "San Francisco"},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "New York"},
+					},
+				},
+			},
+		},
+		{
+			name:             "thinking then tool call",
+			input:            "I should check the weather...</think>\n<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "I should check the weather...",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "Paris"},
+					},
+				},
+			},
+		},
+		{
+			name:             "thinking content then tool call",
+			input:            "Let me think...</think>\nI'll check for you.\n<tool_call>\n<function=search>\n<parameter=query>\ntest\n</parameter>\n</function>\n</tool_call>",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Let me think...",
+			expectedContent:  "I'll check for you.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "search",
+						Arguments: map[string]any{"query": "test"},
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with multiline parameter value",
+			input:      "<tool_call>\n<function=create_note>\n<parameter=content>\nLine 1\nLine 2\nLine 3\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "create_note",
+						Arguments: map[string]any{"content": "Line 1\nLine 2\nLine 3"},
+					},
+				},
+			},
+		},
+		{
+			name:             "empty thinking block - immediate close",
+			input:            "</think>\nHere is my answer.",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "",
+			expectedContent:  "Here is my answer.",
+		},
+		{
+			name:            "thinking disabled but model outputs think close anyway",
+			input:           "</think>\nSome content after spurious tag.",
+			thinkValue:      &api.ThinkValue{Value: false},
+			expectedContent: "</think>\nSome content after spurious tag.",
+		},
+		{
+			name:          "tool call with no function name - returns empty tool call",
+			input:         "<tool_call>\n<function=>\n</function>\n</tool_call>",
+			thinkValue:    nil,
+			expectedCalls: []api.ToolCall{{Function: api.ToolCallFunction{Name: "", Arguments: nil}}},
+		},
+		{
+			name:            "content with newlines preserved",
+			input:           "Line 1\n\nLine 2\n\n\nLine 3",
+			thinkValue:      nil,
+			expectedContent: "Line 1\n\nLine 2\n\n\nLine 3",
+		},
+		{
+			name:             "thinking with only whitespace after close tag",
+			input:            "My thoughts...</think>   \n\t\n   Content here.",
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "My thoughts...",
+			expectedContent:  "Content here.",
+		},
+		{
+			name:            "unicode content",
+			input:           "Hello 世界! 🌍 Ñoño",
+			thinkValue:      nil,
+			expectedContent: "Hello 世界! 🌍 Ñoño",
+		},
+		{
+			name:       "tool call with numeric parameter",
+			input:      "<tool_call>\n<function=set_temp>\n<parameter=value>\n42\n</parameter>\n</function>\n</tool_call>",
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "set_temp",
+						Arguments: map[string]any{"value": "42"},
+					},
+				},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			p := &Nemotron3NanoParser{}
+			p.Init(nil, nil, tt.thinkValue)
+
+			content, thinking, calls, err := p.Add(tt.input, false)
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+
+			// Drain remaining content
+			finalContent, finalThinking, finalCalls, err := p.Add("", true)
+			if err != nil {
+				t.Fatalf("unexpected error on done: %v", err)
+			}
+			content += finalContent
+			thinking += finalThinking
+			calls = append(calls, finalCalls...)
+
+			if diff := cmp.Diff(content, tt.expectedContent); diff != "" {
+				t.Errorf("content mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(thinking, tt.expectedThinking); diff != "" {
+				t.Errorf("thinking mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(calls, tt.expectedCalls); diff != "" {
+				t.Errorf("calls mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestNemotron3NanoParser_Streaming(t *testing.T) {
+	tests := []struct {
+		name             string
+		chunks           []string
+		thinkValue       *api.ThinkValue
+		expectedContent  string
+		expectedThinking string
+		expectedCalls    []api.ToolCall
+	}{
+		{
+			name:            "streaming content character by character",
+			chunks:          []string{"H", "e", "l", "l", "o", ",", " ", "w", "o", "r", "l", "d", "!"},
+			thinkValue:      nil,
+			expectedContent: "Hello, world!",
+		},
+		{
+			name:            "streaming content small tokens",
+			chunks:          []string{"Hel", "lo", ", ", "how ", "can", " I", " help", " you", " today", "?"},
+			thinkValue:      nil,
+			expectedContent: "Hello, how can I help you today?",
+		},
+		{
+			name:             "streaming thinking then content - granular",
+			chunks:           []string{"Let", " me", " th", "ink", " about", " this", "...", "<", "/", "think", ">", "\n", "Here", " is", " my", " answer", "."},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Let me think about this...",
+			expectedContent:  "Here is my answer.",
+		},
+		{
+			name:             "streaming thinking with newlines - granular",
+			chunks:           []string{"Step", " 1", ":", " Ana", "lyze\n", "Step", " 2", ":", " Pro", "cess", "</", "thi", "nk>", "\n", "The", " ans", "wer."},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Step 1: Analyze\nStep 2: Process",
+			expectedContent:  "The answer.",
+		},
+		{
+			name:       "streaming tool call - highly granular",
+			chunks:     []string{"<", "tool", "_", "call", ">", "\n", "<", "func", "tion", "=", "get", "_", "weather", ">", "\n", "<", "param", "eter", "=", "city", ">", "\n", "Par", "is", "\n", "</", "param", "eter", ">", "\n", "</", "func", "tion", ">", "\n", "</", "tool", "_", "call", ">"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "Paris"},
+					},
+				},
+			},
+		},
+		{
+			name:            "streaming content then tool call - granular",
+			chunks:          []string{"Let", " me", " check", " the", " weather", ".", "\n<", "tool_call", ">", "\n", "<function=", "get_weather", ">", "\n", "<parameter=", "city", ">", "\n", "NYC", "\n", "</parameter>", "\n", "</function>", "\n", "</tool_call>"},
+			thinkValue:      nil,
+			expectedContent: "Let me check the weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "NYC"},
+					},
+				},
+			},
+		},
+		{
+			name:   "tool call tag split character by character",
+			chunks: []string{"<", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">", "\n", "<", "f", "u", "n", "c", "t", "i", "o", "n", "=", "t", "e", "s", "t", ">", "\n", "<", "/", "f", "u", "n", "c", "t", "i", "o", "n", ">", "\n", "<", "/", "t", "o", "o", "l", "_", "c", "a", "l", "l", ">"},
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: map[string]any{},
+					},
+				},
+			},
+		},
+		{
+			name:             "thinking close tag split character by character",
+			chunks:           []string{"I", "'", "m", " ", "t", "h", "i", "n", "k", "i", "n", "g", ".", ".", ".", "<", "/", "t", "h", "i", "n", "k", ">", "\n", "D", "o", "n", "e", "!"},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "I'm thinking...",
+			expectedContent:  "Done!",
+		},
+		{
+			name:             "multiple whitespace after think tag - separate chunks",
+			chunks:           []string{"Thinking...", "</think>", "\n", "\n", " ", "Content here."},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Thinking...",
+			expectedContent:  "Content here.",
+		},
+		{
+			name:       "tool call with multiple parameters - streaming",
+			chunks:     []string{"<tool_", "call>\n", "<function", "=book_", "flight>", "\n<para", "meter=", "from>\n", "SFO\n", "</param", "eter>", "\n<param", "eter=to", ">\nNYC", "\n</para", "meter>", "\n</func", "tion>\n", "</tool_", "call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "book_flight",
+						Arguments: map[string]any{
+							"from": "SFO",
+							"to":   "NYC",
+						},
+					},
+				},
+			},
+		},
+		{
+			name:             "thinking then content then tool call - streaming",
+			chunks:           []string{"Ana", "lyzing", " your", " request", "...", "</", "think", ">\n", "I'll", " check", " that", " for", " you", ".", "\n", "<tool", "_call", ">\n", "<function", "=search", ">\n", "<parameter", "=query", ">\n", "test", " query", "\n</", "parameter", ">\n", "</function", ">\n", "</tool", "_call", ">"},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Analyzing your request...",
+			expectedContent:  "I'll check that for you.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "search",
+						Arguments: map[string]any{"query": "test query"},
+					},
+				},
+			},
+		},
+		{
+			name: "multiple tool calls - streaming",
+			chunks: []string{
+				"<tool_call>", "\n", "<function=", "get_weather>", "\n",
+				"<parameter=", "city>\n", "San Fran", "cisco\n", "</parameter>", "\n",
+				"</function>", "\n", "</tool_call>", "\n",
+				"<tool_", "call>\n", "<function", "=get_weather", ">\n",
+				"<param", "eter=city", ">\nNew", " York\n", "</parameter>\n",
+				"</function>\n", "</tool_call>",
+			},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "San Francisco"},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"city": "New York"},
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with multiline parameter - streaming",
+			chunks:     []string{"<tool_call>\n", "<function=", "create_note>\n", "<parameter=", "content>\n", "Line 1", "\nLine", " 2\n", "Line 3", "\n</parameter>\n", "</function>\n", "</tool_call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "create_note",
+						Arguments: map[string]any{"content": "Line 1\nLine 2\nLine 3"},
+					},
+				},
+			},
+		},
+		{
+			name:             "empty thinking block",
+			chunks:           []string{"</think>", "\n", "Just content."},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "",
+			expectedContent:  "Just content.",
+		},
+		{
+			name:            "empty input chunks interspersed",
+			chunks:          []string{"Hello", "", " ", "", "world", "", "!"},
+			thinkValue:      nil,
+			expectedContent: "Hello world!",
+		},
+		{
+			name:             "tool call immediately after think close - no content",
+			chunks:           []string{"Analyzing...", "</think>", "\n", "<tool_call>", "\n<function=test>\n</function>\n", "</tool_call>"},
+			thinkValue:       &api.ThinkValue{Value: true},
+			expectedThinking: "Analyzing...",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: map[string]any{},
+					},
+				},
+			},
+		},
+		{
+			name:       "tool call with empty parameter value",
+			chunks:     []string{"<tool_call>\n<function=test>\n<parameter=name>\n", "\n</parameter>\n</function>\n</tool_call>"},
+			thinkValue: nil,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: map[string]any{"name": ""},
+					},
+				},
+			},
+		},
+		{
+			name:            "partial tool call tag at end - buffered",
+			chunks:          []string{"Here's some content", "<tool"},
+			thinkValue:      nil,
+			expectedContent: "Here's some content",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			p := &Nemotron3NanoParser{}
+			p.Init(nil, nil, tt.thinkValue)
+
+			var allContent string
+			var allThinking string
+			var allCalls []api.ToolCall
+
+			for _, chunk := range tt.chunks {
+				content, thinking, calls, err := p.Add(chunk, false)
+				if err != nil {
+					t.Fatalf("unexpected error: %v", err)
+				}
+				allContent += content
+				allThinking += thinking
+				allCalls = append(allCalls, calls...)
+			}
+
+			// Drain
+			content, thinking, calls, err := p.Add("", true)
+			if err != nil {
+				t.Fatalf("unexpected error on done: %v", err)
+			}
+			allContent += content
+			allThinking += thinking
+			allCalls = append(allCalls, calls...)
+
+			if diff := cmp.Diff(allContent, tt.expectedContent); diff != "" {
+				t.Errorf("content mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(allThinking, tt.expectedThinking); diff != "" {
+				t.Errorf("thinking mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(allCalls, tt.expectedCalls); diff != "" {
+				t.Errorf("calls mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestNemotron3NanoParser_HasToolSupport(t *testing.T) {
+	p := &Nemotron3NanoParser{}
+	if !p.HasToolSupport() {
+		t.Error("expected HasToolSupport to return true")
+	}
+}
+
+func TestNemotron3NanoParser_HasThinkingSupport(t *testing.T) {
+	p := &Nemotron3NanoParser{}
+	if !p.HasThinkingSupport() {
+		t.Error("expected HasThinkingSupport to return true")
+	}
+}
+
+func TestNemotron3NanoParser_Init(t *testing.T) {
+	t.Run("starts in thinking state when enabled", func(t *testing.T) {
+		p := &Nemotron3NanoParser{}
+		p.Init(nil, nil, &api.ThinkValue{Value: true})
+		if p.state != Nemotron3NanoCollectingThinking {
+			t.Errorf("expected state Nemotron3NanoCollectingThinking, got %v", p.state)
+		}
+	})
+
+	t.Run("starts in content state when thinking disabled", func(t *testing.T) {
+		p := &Nemotron3NanoParser{}
+		p.Init(nil, nil, &api.ThinkValue{Value: false})
+		if p.state != Nemotron3NanoCollectingContent {
+			t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
+		}
+	})
+
+	t.Run("starts in content state when nil thinkValue", func(t *testing.T) {
+		p := &Nemotron3NanoParser{}
+		p.Init(nil, nil, nil)
+		if p.state != Nemotron3NanoCollectingContent {
+			t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
+		}
+	})
+
+	t.Run("starts in content state with assistant prefill", func(t *testing.T) {
+		p := &Nemotron3NanoParser{}
+		prefill := &api.Message{Role: "assistant", Content: "Starting..."}
+		p.Init(nil, prefill, &api.ThinkValue{Value: true})
+		if p.state != Nemotron3NanoCollectingContent {
+			t.Errorf("expected state Nemotron3NanoCollectingContent, got %v", p.state)
+		}
+	})
+}
+
+func TestNemotron3NanoParser_WithTools(t *testing.T) {
+	tools := []api.Tool{
+		{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name: "get_weather",
+				Parameters: api.ToolFunctionParameters{
+					Type: "object",
+					Properties: map[string]api.ToolProperty{
+						"city": {Type: api.PropertyType{"string"}},
+					},
+				},
+			},
+		},
+	}
+
+	p := &Nemotron3NanoParser{}
+	returnedTools := p.Init(tools, nil, nil)
+
+	if diff := cmp.Diff(returnedTools, tools); diff != "" {
+		t.Errorf("tools mismatch (-got +want):\n%s", diff)
+	}
+
+	// Parse a tool call
+	input := "<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>"
+	_, _, calls, err := p.Add(input, true)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	expectedCalls := []api.ToolCall{
+		{
+			Function: api.ToolCallFunction{
+				Name:      "get_weather",
+				Arguments: map[string]any{"city": "Paris"},
+			},
+		},
+	}
+
+	if diff := cmp.Diff(calls, expectedCalls); diff != "" {
+		t.Errorf("calls mismatch (-got +want):\n%s", diff)
+	}
+}

model/parsers/parsers.go

@@ -62,6 +62,8 @@ func ParserForName(name string) Parser {
 		return &Olmo3Parser{}
 	case "olmo3-think":
 		return &Olmo3ThinkParser{}
+	case "nemotron-3-nano":
+		return &Nemotron3NanoParser{}
 	default:
 		return nil
 	}

model/renderers/deepseek3.go

@@ -0,0 +1,146 @@
+package renderers
+
+import (
+	"encoding/json"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+type DeepSeek3Variant int
+
+const (
+	Deepseek31 DeepSeek3Variant = iota
+)
+
+type DeepSeek3Renderer struct {
+	IsThinking bool
+	Variant    DeepSeek3Variant
+}
+
+func (r *DeepSeek3Renderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
+	var sb strings.Builder
+
+	// thinking is enabled: model must support it AND user must request it
+	thinking := r.IsThinking && (thinkValue != nil && thinkValue.Bool())
+
+	// extract system messages first
+	var systemPrompt strings.Builder
+	isFirstSystemPrompt := true
+
+	for _, message := range messages {
+		if message.Role == "system" {
+			if isFirstSystemPrompt {
+				systemPrompt.WriteString(message.Content)
+				isFirstSystemPrompt = false
+			} else {
+				systemPrompt.WriteString("\n\n" + message.Content)
+			}
+		}
+	}
+
+	sb.WriteString("<｜begin▁of▁sentence｜>")
+	sb.WriteString(systemPrompt.String())
+
+	// tool definitions
+	if len(tools) > 0 {
+		sb.WriteString("\n\n## Tools\nYou have access to the following tools:\n")
+
+		for _, tool := range tools {
+			sb.WriteString("\n### " + tool.Function.Name)
+			sb.WriteString("\nDescription: " + tool.Function.Description)
+
+			// parameters as JSON
+			parametersJSON, err := json.Marshal(tool.Function.Parameters)
+			if err == nil {
+				sb.WriteString("\n\nParameters: " + string(parametersJSON) + "\n")
+			}
+		}
+
+		// usage instructions
+		sb.WriteString("\nIMPORTANT: ALWAYS adhere to this exact format for tool use:\n")
+		sb.WriteString("<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>\n\n")
+		sb.WriteString("Where:\n\n")
+		sb.WriteString("- `tool_call_name` must be an exact match to one of the available tools\n")
+		sb.WriteString("- `tool_call_arguments` must be valid JSON that strictly follows the tool's Parameters Schema\n")
+		sb.WriteString("- For multiple tool calls, chain them directly without separators or spaces\n")
+	}
+
+	// state tracking
+	isTool := false
+	isLastUser := false
+
+	for _, message := range messages {
+		switch message.Role {
+		case "user":
+			isTool = false
+			isLastUser = true
+			sb.WriteString("<｜User｜>" + message.Content)
+
+		case "assistant":
+			if len(message.ToolCalls) > 0 {
+				if isLastUser {
+					sb.WriteString("<｜Assistant｜></think>")
+				}
+				isLastUser = false
+				isTool = false
+
+				if message.Content != "" {
+					sb.WriteString(message.Content)
+				}
+
+				sb.WriteString("<｜tool▁calls▁begin｜>")
+				for _, toolCall := range message.ToolCalls {
+					sb.WriteString("<｜tool▁call▁begin｜>" + toolCall.Function.Name + "<｜tool▁sep｜>")
+
+					argsJSON, _ := json.Marshal(toolCall.Function.Arguments)
+					sb.WriteString(string(argsJSON))
+					sb.WriteString("<｜tool▁call▁end｜>")
+				}
+				sb.WriteString("<｜tool▁calls▁end｜><｜end▁of▁sentence｜>")
+			} else {
+				if isLastUser {
+					sb.WriteString("<｜Assistant｜>")
+					// message["prefix"] is defined and message["prefix"] and thinking
+					// message.Thinking != "" represents message["prefix"] being defined
+					if message.Thinking != "" && thinking {
+						sb.WriteString("<think>")
+					} else {
+						sb.WriteString("</think>")
+					}
+				}
+				isLastUser = false
+
+				content := message.Content
+				if isTool {
+					sb.WriteString(content + "<｜end▁of▁sentence｜>")
+					isTool = false
+				} else {
+					if strings.Contains(content, "</think>") {
+						parts := strings.SplitN(content, "</think>", 2)
+						if len(parts) > 1 {
+							content = parts[1]
+						}
+					}
+					sb.WriteString(content + "<｜end▁of▁sentence｜>")
+				}
+			}
+
+		case "tool":
+			isLastUser = false
+			isTool = true
+			sb.WriteString("<｜tool▁output▁begin｜>" + message.Content + "<｜tool▁output▁end｜>")
+		}
+	}
+
+	if isLastUser && !isTool {
+		sb.WriteString("<｜Assistant｜>")
+		if thinking {
+			sb.WriteString("<think>")
+		} else {
+			sb.WriteString("</think>")
+		}
+	}
+
+	return sb.String(), nil
+}

model/renderers/deepseek3_test.go

@@ -0,0 +1,976 @@
+package renderers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestDeepSeekRenderer(t *testing.T) {
+	tests := []struct {
+		name       string
+		messages   []api.Message
+		tools      []api.Tool
+		thinkValue *api.ThinkValue
+		expected   string
+	}{
+		{
+			name: "basic user message",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello, how are you?"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Hello, how are you?<｜Assistant｜></think>`,
+		},
+		{
+			name: "basic with system message",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello, how are you?"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜>You are a helpful assistant.<｜User｜>Hello, how are you?<｜Assistant｜></think>`,
+		},
+		{
+			name: "multiple system messages",
+			messages: []api.Message{
+				{Role: "system", Content: "First instruction"},
+				{Role: "system", Content: "Second instruction"},
+				{Role: "user", Content: "Hello"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected: `<｜begin▁of▁sentence｜>First instruction
+
+Second instruction<｜User｜>Hello<｜Assistant｜></think>`,
+		},
+		{
+			name: "thinking enabled",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello, how are you?"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Hello, how are you?<｜Assistant｜><think>`,
+		},
+		{
+			name: "thinking enabled with system",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello, how are you?"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜>You are a helpful assistant.<｜User｜>Hello, how are you?<｜Assistant｜><think>`,
+		},
+		{
+			name: "conversation with assistant response",
+			messages: []api.Message{
+				{Role: "user", Content: "What is the capital of France?"},
+				{Role: "assistant", Content: "The capital of France is Paris."},
+				{Role: "user", Content: "Fantastic!"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>What is the capital of France?<｜Assistant｜></think>The capital of France is Paris.<｜end▁of▁sentence｜><｜User｜>Fantastic!<｜Assistant｜></think>`,
+		},
+		{
+			name: "assistant with tool calls",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>What's the weather?<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "assistant with content and tool calls",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather in Paris?"},
+				{
+					Role:    "assistant",
+					Content: "I'll check the weather for you.",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>What's the weather in Paris?<｜Assistant｜></think>I'll check the weather for you.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "tool response",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Temperature: 22°C, Sunny"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>What's the weather?<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Temperature: 22°C, Sunny<｜tool▁output▁end｜>`,
+		},
+		{
+			name: "multiple tool calls",
+			messages: []api.Message{
+				{Role: "user", Content: "Get weather for Paris and London"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "London",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Paris: 22°C, Sunny"},
+				{Role: "tool", Content: "London: 18°C, Cloudy"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Get weather for Paris and London<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"London"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Paris: 22°C, Sunny<｜tool▁output▁end｜><｜tool▁output▁begin｜>London: 18°C, Cloudy<｜tool▁output▁end｜>`,
+		},
+		{
+			name: "content with </think> tag removal",
+			messages: []api.Message{
+				{Role: "user", Content: "Think about this"},
+				{Role: "assistant", Content: "I'm thinking about this.</think>The answer is 42."},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Think about this<｜Assistant｜></think>The answer is 42.<｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "empty system message",
+			messages: []api.Message{
+				{Role: "system", Content: ""},
+				{Role: "user", Content: "Hello"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Hello<｜Assistant｜></think>`,
+		},
+		{
+			name: "empty assistant content",
+			messages: []api.Message{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: ""},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Hello<｜Assistant｜></think><｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "special characters",
+			messages: []api.Message{
+				{Role: "user", Content: "What about <|special|> tokens and \"quotes\"?"},
+				{Role: "assistant", Content: "They're handled normally."},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>What about <|special|> tokens and "quotes"?<｜Assistant｜></think>They're handled normally.<｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "tool calls with null content",
+			messages: []api.Message{
+				{Role: "user", Content: "Get weather"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Get weather<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "assistant after tool context",
+			messages: []api.Message{
+				{Role: "user", Content: "Process data"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "process",
+								Arguments: api.ToolCallFunctionArguments{
+									"data": "test",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Success"},
+				{Role: "assistant", Content: "Done"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Process data<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>process<｜tool▁sep｜>{"data":"test"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Success<｜tool▁output▁end｜>Done<｜end▁of▁sentence｜>`,
+		},
+		{
+			name:       "no messages",
+			messages:   []api.Message{},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜>`,
+		},
+		{
+			name: "only system messages",
+			messages: []api.Message{
+				{Role: "system", Content: "System instruction"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜>System instruction`,
+		},
+		{
+			name: "multiple think tags in content",
+			messages: []api.Message{
+				{Role: "user", Content: "Complex question"},
+				{Role: "assistant", Content: "First thought</think>Second thought</think>Final answer"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Complex question<｜Assistant｜></think>Second thought</think>Final answer<｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "thinking enabled after tool call - should render thinking",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather in Paris?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Temperature: 22°C, Sunny"},
+				{Role: "assistant", Content: "Based on the weather data, it's sunny in Paris."},
+				{Role: "user", Content: "Now tell me about London weather too."},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>What's the weather in Paris?<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Temperature: 22°C, Sunny<｜tool▁output▁end｜>Based on the weather data, it's sunny in Paris.<｜end▁of▁sentence｜><｜User｜>Now tell me about London weather too.<｜Assistant｜><think>`,
+		},
+		{
+			name: "thinking disabled after tool call - should not render thinking",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather in Paris?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Temperature: 22°C, Sunny"},
+				{Role: "assistant", Content: "Based on the weather data, it's sunny in Paris."},
+				{Role: "user", Content: "Now tell me about London weather too."},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>What's the weather in Paris?<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Temperature: 22°C, Sunny<｜tool▁output▁end｜>Based on the weather data, it's sunny in Paris.<｜end▁of▁sentence｜><｜User｜>Now tell me about London weather too.<｜Assistant｜></think>`,
+		},
+		{
+			name: "thinking enabled but messages without thinking content",
+			messages: []api.Message{
+				{Role: "user", Content: "First question about cats"},
+				{Role: "assistant", Content: "Cats are wonderful pets."},
+				{Role: "user", Content: "What about dogs?"},
+				{Role: "assistant", Content: "Dogs are loyal companions."},
+				{Role: "user", Content: "Final question about birds"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>First question about cats<｜Assistant｜></think>Cats are wonderful pets.<｜end▁of▁sentence｜><｜User｜>What about dogs?<｜Assistant｜></think>Dogs are loyal companions.<｜end▁of▁sentence｜><｜User｜>Final question about birds<｜Assistant｜><think>`,
+		},
+		{
+			name: "thinking disabled for all assistant responses",
+			messages: []api.Message{
+				{Role: "user", Content: "First question about cats"},
+				{Role: "assistant", Content: "Cats are wonderful pets."},
+				{Role: "user", Content: "What about dogs?"},
+				{Role: "assistant", Content: "Dogs are loyal companions."},
+				{Role: "user", Content: "Final question about birds"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>First question about cats<｜Assistant｜></think>Cats are wonderful pets.<｜end▁of▁sentence｜><｜User｜>What about dogs?<｜Assistant｜></think>Dogs are loyal companions.<｜end▁of▁sentence｜><｜User｜>Final question about birds<｜Assistant｜></think>`,
+		},
+		{
+			name: "complex conversation with tool calls and thinking enabled",
+			messages: []api.Message{
+				{Role: "user", Content: "Tell me about the weather"},
+				{Role: "assistant", Content: "I'll check the weather for you."},
+				{Role: "user", Content: "Actually, get Paris weather specifically"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Paris: 22°C, Sunny"},
+				{Role: "assistant", Content: "The weather in Paris is great!"},
+				{Role: "user", Content: "What about the forecast for tomorrow?"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Tell me about the weather<｜Assistant｜></think>I'll check the weather for you.<｜end▁of▁sentence｜><｜User｜>Actually, get Paris weather specifically<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Paris: 22°C, Sunny<｜tool▁output▁end｜>The weather in Paris is great!<｜end▁of▁sentence｜><｜User｜>What about the forecast for tomorrow?<｜Assistant｜><think>`,
+		},
+		{
+			name: "tool call without subsequent user message - no thinking",
+			messages: []api.Message{
+				{Role: "user", Content: "Get the weather"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "22°C, Sunny"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Get the weather<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>22°C, Sunny<｜tool▁output▁end｜>`,
+		},
+		{
+			name: "messages with thinking content, no thinking in render",
+			messages: []api.Message{
+				{Role: "user", Content: "Solve this math problem: 15 * 23"},
+				{
+					Role:     "assistant",
+					Content:  "The answer is 345.",
+					Thinking: "Let me calculate 15 * 23. I can break this down: 15 * 20 = 300, and 15 * 3 = 45, so 300 + 45 = 345.",
+				},
+				{Role: "user", Content: "What about 12 * 34?"},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Solve this math problem: 15 * 23<｜Assistant｜></think>The answer is 345.<｜end▁of▁sentence｜><｜User｜>What about 12 * 34?<｜Assistant｜></think>`,
+		},
+		{
+			name: "conversation with mix of thinking and no thinking",
+			messages: []api.Message{
+				{Role: "user", Content: "Explain quantum physics"},
+				{
+					Role:     "assistant",
+					Content:  "Quantum physics is the study of matter and energy at the smallest scales.",
+					Thinking: "This is a complex topic. I should start with basic concepts and avoid overwhelming technical details.",
+				},
+				{Role: "user", Content: "What about photons?"},
+				{
+					Role:    "assistant",
+					Content: "Photons are particles of light with no mass.",
+				},
+				{Role: "user", Content: "How do they interact with matter?"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Explain quantum physics<｜Assistant｜><think>Quantum physics is the study of matter and energy at the smallest scales.<｜end▁of▁sentence｜><｜User｜>What about photons?<｜Assistant｜></think>Photons are particles of light with no mass.<｜end▁of▁sentence｜><｜User｜>How do they interact with matter?<｜Assistant｜><think>`,
+		},
+		{
+			name: "tool call with thinking content in response",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather in Tokyo and New York?"},
+				{
+					Role:     "assistant",
+					Content:  "I'll check the weather for both cities.",
+					Thinking: "I need to call the weather API for two different cities. Let me make parallel calls.",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Tokyo",
+								},
+							},
+						},
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "New York",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Tokyo: 18°C, Cloudy"},
+				{Role: "tool", Content: "New York: 22°C, Sunny"},
+				{
+					Role:     "assistant",
+					Content:  "Based on the weather data: Tokyo is cloudy at 18°C, while New York is sunny at 22°C.",
+					Thinking: "The data shows a nice contrast between the two cities. Tokyo is cooler and overcast while NYC has better weather.",
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>What's the weather in Tokyo and New York?<｜Assistant｜></think>I'll check the weather for both cities.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Tokyo"}<｜tool▁call▁end｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"New York"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Tokyo: 18°C, Cloudy<｜tool▁output▁end｜><｜tool▁output▁begin｜>New York: 22°C, Sunny<｜tool▁output▁end｜>Based on the weather data: Tokyo is cloudy at 18°C, while New York is sunny at 22°C.<｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "empty thinking field",
+			messages: []api.Message{
+				{Role: "user", Content: "Simple question"},
+				{
+					Role:     "assistant",
+					Content:  "Simple answer.",
+					Thinking: "", // Empty thinking content
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected:   `<｜begin▁of▁sentence｜><｜User｜>Simple question<｜Assistant｜></think>Simple answer.<｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "with tools definitions",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "What's the weather like?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather information",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							},
+							Required: []string{"location"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected: `<｜begin▁of▁sentence｜>You are a helpful assistant.
+
+## Tools
+You have access to the following tools:
+
+### get_weather
+Description: Get current weather information
+
+Parameters: {"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}
+
+IMPORTANT: ALWAYS adhere to this exact format for tool use:
+<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>
+
+Where:
+
+- ` + "`tool_call_name`" + ` must be an exact match to one of the available tools
+- ` + "`tool_call_arguments`" + ` must be valid JSON that strictly follows the tool's Parameters Schema
+- For multiple tool calls, chain them directly without separators or spaces
+<｜User｜>What's the weather like?<｜Assistant｜></think>`,
+		},
+		{
+			name: "tools definitions with thinking enabled",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "What's the weather in Paris?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather information",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							},
+							Required: []string{"location"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: `<｜begin▁of▁sentence｜>You are a helpful assistant.
+
+## Tools
+You have access to the following tools:
+
+### get_weather
+Description: Get current weather information
+
+Parameters: {"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}
+
+IMPORTANT: ALWAYS adhere to this exact format for tool use:
+<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>
+
+Where:
+
+- ` + "`tool_call_name`" + ` must be an exact match to one of the available tools
+- ` + "`tool_call_arguments`" + ` must be valid JSON that strictly follows the tool's Parameters Schema
+- For multiple tool calls, chain them directly without separators or spaces
+<｜User｜>What's the weather in Paris?<｜Assistant｜><think>`,
+		},
+		{
+			name: "tools definitions with actual tool call",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "What's the weather in Paris?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather information",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							},
+							Required: []string{"location"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected: `<｜begin▁of▁sentence｜>You are a helpful assistant.
+
+## Tools
+You have access to the following tools:
+
+### get_weather
+Description: Get current weather information
+
+Parameters: {"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}
+
+IMPORTANT: ALWAYS adhere to this exact format for tool use:
+<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>
+
+Where:
+
+- ` + "`tool_call_name`" + ` must be an exact match to one of the available tools
+- ` + "`tool_call_arguments`" + ` must be valid JSON that strictly follows the tool's Parameters Schema
+- For multiple tool calls, chain them directly without separators or spaces
+<｜User｜>What's the weather in Paris?<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "tools definitions with full conversation cycle",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "What's the weather in Paris?"},
+				{
+					Role:    "assistant",
+					Content: "I'll check the weather for you.",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Temperature: 22°C, Sunny"},
+				{Role: "assistant", Content: "The weather in Paris is 22°C and sunny!"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather information",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							},
+							Required: []string{"location"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected: `<｜begin▁of▁sentence｜>You are a helpful assistant.
+
+## Tools
+You have access to the following tools:
+
+### get_weather
+Description: Get current weather information
+
+Parameters: {"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}
+
+IMPORTANT: ALWAYS adhere to this exact format for tool use:
+<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>
+
+Where:
+
+- ` + "`tool_call_name`" + ` must be an exact match to one of the available tools
+- ` + "`tool_call_arguments`" + ` must be valid JSON that strictly follows the tool's Parameters Schema
+- For multiple tool calls, chain them directly without separators or spaces
+<｜User｜>What's the weather in Paris?<｜Assistant｜></think>I'll check the weather for you.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Temperature: 22°C, Sunny<｜tool▁output▁end｜>The weather in Paris is 22°C and sunny!<｜end▁of▁sentence｜>`,
+		},
+		{
+			name: "tools with thinking and full conversation",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Check the weather in Tokyo"},
+				{
+					Role:     "assistant",
+					Thinking: "The user wants weather info for Tokyo. I should use the get_weather tool.",
+					Content:  "Let me check that for you.",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Tokyo",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Temperature: 18°C, Cloudy"},
+				{
+					Role:     "assistant",
+					Thinking: "The weather data shows it's cloudy and cool. I should present this clearly.",
+					Content:  "In Tokyo, it's currently 18°C and cloudy.",
+				},
+				{Role: "user", Content: "What about London?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather information",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							},
+							Required: []string{"location"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: `<｜begin▁of▁sentence｜>You are a helpful assistant.
+
+## Tools
+You have access to the following tools:
+
+### get_weather
+Description: Get current weather information
+
+Parameters: {"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}
+
+IMPORTANT: ALWAYS adhere to this exact format for tool use:
+<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>
+
+Where:
+
+- ` + "`tool_call_name`" + ` must be an exact match to one of the available tools
+- ` + "`tool_call_arguments`" + ` must be valid JSON that strictly follows the tool's Parameters Schema
+- For multiple tool calls, chain them directly without separators or spaces
+<｜User｜>Check the weather in Tokyo<｜Assistant｜></think>Let me check that for you.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Tokyo"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Temperature: 18°C, Cloudy<｜tool▁output▁end｜>In Tokyo, it's currently 18°C and cloudy.<｜end▁of▁sentence｜><｜User｜>What about London?<｜Assistant｜><think>`,
+		},
+		{
+			name: "multiple tools definitions",
+			messages: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant with access to multiple tools."},
+				{Role: "user", Content: "What can you help me with?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather information",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							},
+							Required: []string{"location"},
+						},
+					},
+				},
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "calculate",
+						Description: "Perform mathematical calculations",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"expression": {
+									Type:        api.PropertyType{"string"},
+									Description: "Mathematical expression to evaluate",
+								},
+							},
+							Required: []string{"expression"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected: `<｜begin▁of▁sentence｜>You are a helpful assistant with access to multiple tools.
+
+## Tools
+You have access to the following tools:
+
+### get_weather
+Description: Get current weather information
+
+Parameters: {"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}
+
+### calculate
+Description: Perform mathematical calculations
+
+Parameters: {"type":"object","required":["expression"],"properties":{"expression":{"type":"string","description":"Mathematical expression to evaluate"}}}
+
+IMPORTANT: ALWAYS adhere to this exact format for tool use:
+<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>
+
+Where:
+
+- ` + "`tool_call_name`" + ` must be an exact match to one of the available tools
+- ` + "`tool_call_arguments`" + ` must be valid JSON that strictly follows the tool's Parameters Schema
+- For multiple tool calls, chain them directly without separators or spaces
+<｜User｜>What can you help me with?<｜Assistant｜></think>`,
+		},
+		{
+			name: "multiple tools with multiple tool calls",
+			messages: []api.Message{
+				{Role: "user", Content: "Get weather for Paris and calculate 25 * 4"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: api.ToolCallFunctionArguments{
+									"location": "Paris",
+								},
+							},
+						},
+						{
+							Function: api.ToolCallFunction{
+								Name: "calculate",
+								Arguments: api.ToolCallFunctionArguments{
+									"expression": "25 * 4",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Temperature: 22°C, Sunny"},
+				{Role: "tool", Content: "Result: 100"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather information",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							},
+							Required: []string{"location"},
+						},
+					},
+				},
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "calculate",
+						Description: "Perform mathematical calculations",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"expression": {
+									Type:        api.PropertyType{"string"},
+									Description: "Mathematical expression to evaluate",
+								},
+							},
+							Required: []string{"expression"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected: `<｜begin▁of▁sentence｜>
+
+## Tools
+You have access to the following tools:
+
+### get_weather
+Description: Get current weather information
+
+Parameters: {"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}
+
+### calculate
+Description: Perform mathematical calculations
+
+Parameters: {"type":"object","required":["expression"],"properties":{"expression":{"type":"string","description":"Mathematical expression to evaluate"}}}
+
+IMPORTANT: ALWAYS adhere to this exact format for tool use:
+<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>
+
+Where:
+
+- ` + "`tool_call_name`" + ` must be an exact match to one of the available tools
+- ` + "`tool_call_arguments`" + ` must be valid JSON that strictly follows the tool's Parameters Schema
+- For multiple tool calls, chain them directly without separators or spaces
+<｜User｜>Get weather for Paris and calculate 25 * 4<｜Assistant｜></think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{"location":"Paris"}<｜tool▁call▁end｜><｜tool▁call▁begin｜>calculate<｜tool▁sep｜>{"expression":"25 * 4"}<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜><｜tool▁output▁begin｜>Temperature: 22°C, Sunny<｜tool▁output▁end｜><｜tool▁output▁begin｜>Result: 100<｜tool▁output▁end｜>`,
+		},
+		{
+			name: "tools without system message",
+			messages: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get current weather information",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {
+									Type:        api.PropertyType{"string"},
+									Description: "City name",
+								},
+							},
+							Required: []string{"location"},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: false},
+			expected: `<｜begin▁of▁sentence｜>
+
+## Tools
+You have access to the following tools:
+
+### get_weather
+Description: Get current weather information
+
+Parameters: {"type":"object","required":["location"],"properties":{"location":{"type":"string","description":"City name"}}}
+
+IMPORTANT: ALWAYS adhere to this exact format for tool use:
+<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool_call_name<｜tool▁sep｜>tool_call_arguments<｜tool▁call▁end｜>{{additional_tool_calls}}<｜tool▁calls▁end｜>
+
+Where:
+
+- ` + "`tool_call_name`" + ` must be an exact match to one of the available tools
+- ` + "`tool_call_arguments`" + ` must be valid JSON that strictly follows the tool's Parameters Schema
+- For multiple tool calls, chain them directly without separators or spaces
+<｜User｜>What's the weather?<｜Assistant｜></think>`,
+		},
+	}
+
+	renderer := &DeepSeek3Renderer{IsThinking: true, Variant: Deepseek31}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			rendered, err := renderer.Render(tt.messages, tt.tools, tt.thinkValue)
+			if err != nil {
+				t.Fatalf("Render() error = %v", err)
+			}
+			if diff := cmp.Diff(tt.expected, rendered); diff != "" {
+				t.Errorf("Render() mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}

model/renderers/nemotron3nano.go

@@ -0,0 +1,220 @@
+package renderers
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+type Nemotron3NanoRenderer struct{}
+
+func (r *Nemotron3NanoRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
+	var sb strings.Builder
+
+	// thinking is enabled if user requests it
+	enableThinking := thinkValue != nil && thinkValue.Bool()
+
+	// Extract system message if present
+	var systemMessage string
+	var loopMessages []api.Message
+	if len(messages) > 0 && messages[0].Role == "system" {
+		systemMessage = messages[0].Content
+		loopMessages = messages[1:]
+	} else {
+		loopMessages = messages
+	}
+
+	// Find last user message index for thinking truncation
+	lastUserIdx := -1
+	for i, msg := range loopMessages {
+		if msg.Role == "user" {
+			lastUserIdx = i
+		}
+	}
+
+	sb.WriteString("<|im_start|>system\n")
+	if systemMessage != "" {
+		sb.WriteString(systemMessage)
+	}
+
+	if len(tools) > 0 {
+		if systemMessage != "" {
+			sb.WriteString("\n\n")
+		}
+		sb.WriteString(r.renderTools(tools))
+	}
+	sb.WriteString("<|im_end|>\n")
+
+	for i, message := range loopMessages {
+		switch message.Role {
+		case "assistant":
+			// Build content with thinking tags
+			content := r.buildContent(message)
+			shouldTruncate := i < lastUserIdx
+
+			if len(message.ToolCalls) > 0 {
+				sb.WriteString("<|im_start|>assistant\n")
+				sb.WriteString(r.formatContent(content, shouldTruncate, true))
+				r.writeToolCalls(&sb, message.ToolCalls)
+				sb.WriteString("<|im_end|>\n")
+			} else {
+				formatted := r.formatContent(content, shouldTruncate, false)
+				sb.WriteString("<|im_start|>assistant\n" + formatted + "<|im_end|>\n")
+			}
+
+		case "user", "system":
+			sb.WriteString("<|im_start|>" + message.Role + "\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("<|im_end|>\n")
+
+		case "tool":
+			// Check if previous message was also a tool message
+			prevWasTool := i > 0 && loopMessages[i-1].Role == "tool"
+			nextIsTool := i+1 < len(loopMessages) && loopMessages[i+1].Role == "tool"
+
+			if !prevWasTool {
+				sb.WriteString("<|im_start|>user\n")
+			}
+			sb.WriteString("<tool_response>\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("\n</tool_response>\n")
+
+			if !nextIsTool {
+				sb.WriteString("<|im_end|>\n")
+			}
+
+		default:
+			sb.WriteString("<|im_start|>" + message.Role + "\n" + message.Content + "<|im_end|>\n")
+		}
+	}
+
+	// Add generation prompt
+	if enableThinking {
+		sb.WriteString("<|im_start|>assistant\n<think>\n")
+	} else {
+		sb.WriteString("<|im_start|>assistant\n<think></think>")
+	}
+
+	return sb.String(), nil
+}
+
+func (r *Nemotron3NanoRenderer) renderTools(tools []api.Tool) string {
+	var sb strings.Builder
+	sb.WriteString("# Tools\n\nYou have access to the following functions:\n\n<tools>")
+
+	for _, tool := range tools {
+		fn := tool.Function
+		sb.WriteString("\n<function>\n<name>" + fn.Name + "</name>")
+
+		if fn.Description != "" {
+			sb.WriteString("\n<description>" + strings.TrimSpace(fn.Description) + "</description>")
+		}
+
+		sb.WriteString("\n<parameters>")
+		if fn.Parameters.Properties != nil {
+			for paramName, paramFields := range fn.Parameters.Properties {
+				sb.WriteString("\n<parameter>")
+				sb.WriteString("\n<name>" + paramName + "</name>")
+
+				if len(paramFields.Type) > 0 {
+					sb.WriteString("\n<type>" + strings.Join(paramFields.Type, ", ") + "</type>")
+				}
+
+				if paramFields.Description != "" {
+					sb.WriteString("\n<description>" + strings.TrimSpace(paramFields.Description) + "</description>")
+				}
+
+				if len(paramFields.Enum) > 0 {
+					enumJSON, _ := json.Marshal(paramFields.Enum)
+					sb.WriteString("\n<enum>" + string(enumJSON) + "</enum>")
+				}
+
+				sb.WriteString("\n</parameter>")
+			}
+		}
+
+		if len(fn.Parameters.Required) > 0 {
+			reqJSON, _ := json.Marshal(fn.Parameters.Required)
+			sb.WriteString("\n<required>" + string(reqJSON) + "</required>")
+		}
+
+		sb.WriteString("\n</parameters>")
+		sb.WriteString("\n</function>")
+	}
+
+	sb.WriteString("\n</tools>")
+
+	sb.WriteString("\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+		"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+		"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+		"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+		"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+		"- Required parameters MUST be specified\n" +
+		"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+		"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>")
+
+	return sb.String()
+}
+
+func (r *Nemotron3NanoRenderer) buildContent(message api.Message) string {
+	// The parser always extracts thinking into the Thinking field,
+	// so Content will never have <think> tags embedded
+	if message.Thinking != "" {
+		return "<think>\n" + message.Thinking + "\n</think>\n" + message.Content
+	}
+	return "<think></think>" + message.Content
+}
+
+func (r *Nemotron3NanoRenderer) formatContent(content string, truncate bool, addNewline bool) string {
+	if content == "" {
+		return "<think></think>"
+	}
+
+	if !truncate {
+		if addNewline {
+			return strings.TrimSpace(content) + "\n"
+		}
+		return strings.TrimSpace(content)
+	}
+
+	// Truncate thinking - keep only content after </think>
+	c := content
+	if strings.Contains(c, "</think>") {
+		parts := strings.Split(c, "</think>")
+		c = parts[len(parts)-1]
+	} else if strings.Contains(c, "<think>") {
+		parts := strings.Split(c, "<think>")
+		c = parts[0]
+	}
+	c = "<think></think>" + strings.TrimSpace(c)
+
+	if addNewline && len(c) > len("<think></think>") {
+		return c + "\n"
+	}
+	if c == "<think></think>" {
+		return c
+	}
+	return strings.TrimSpace(c)
+}
+
+func (r *Nemotron3NanoRenderer) writeToolCalls(sb *strings.Builder, toolCalls []api.ToolCall) {
+	for _, tc := range toolCalls {
+		sb.WriteString("<tool_call>\n<function=" + tc.Function.Name + ">\n")
+		for name, value := range tc.Function.Arguments {
+			sb.WriteString("<parameter=" + name + ">\n" + r.formatArgValue(value) + "\n</parameter>\n")
+		}
+		sb.WriteString("</function>\n</tool_call>\n")
+	}
+}
+
+func (r *Nemotron3NanoRenderer) formatArgValue(value any) string {
+	switch v := value.(type) {
+	case map[string]any, []any:
+		jsonBytes, _ := json.Marshal(v)
+		return string(jsonBytes)
+	default:
+		return fmt.Sprintf("%v", v)
+	}
+}

model/renderers/nemotron3nano_test.go

@@ -0,0 +1,569 @@
+package renderers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestNemotron3NanoRenderer(t *testing.T) {
+	tests := []struct {
+		name       string
+		msgs       []api.Message
+		tools      []api.Tool
+		thinkValue *api.ThinkValue
+		expected   string
+	}{
+		{
+			name: "basic user message - thinking mode",
+			msgs: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n<|im_end|>\n" +
+				"<|im_start|>user\nHello!<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "basic user message - no thinking",
+			msgs: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			thinkValue: nil,
+			expected: "<|im_start|>system\n<|im_end|>\n" +
+				"<|im_start|>user\nHello!<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>",
+		},
+		{
+			name: "with system message",
+			msgs: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello!"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n" +
+				"<|im_start|>user\nHello!<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "multi-turn conversation",
+			msgs: []api.Message{
+				{Role: "user", Content: "Hi"},
+				{Role: "assistant", Content: "Hello! How can I help?"},
+				{Role: "user", Content: "Tell me a joke"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n<|im_end|>\n" +
+				"<|im_start|>user\nHi<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>Hello! How can I help?<|im_end|>\n" +
+				"<|im_start|>user\nTell me a joke<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "with tools",
+			msgs: []api.Message{
+				{Role: "user", Content: "What's the weather in Paris?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get the current weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:     "object",
+							Required: []string{"city"},
+							Properties: map[string]api.ToolProperty{
+								"city": {Type: api.PropertyType{"string"}, Description: "The city name"},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n" +
+				"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
+				"<function>\n<name>get_weather</name>\n" +
+				"<description>Get the current weather</description>\n" +
+				"<parameters>\n" +
+				"<parameter>\n<name>city</name>\n<type>string</type>\n<description>The city name</description>\n</parameter>\n" +
+				"<required>[\"city\"]</required>\n" +
+				"</parameters>\n</function>\n</tools>\n\n" +
+				"If you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+				"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+				"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+				"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+				"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+				"- Required parameters MUST be specified\n" +
+				"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+				"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
+				"</IMPORTANT><|im_end|>\n" +
+				"<|im_start|>user\nWhat's the weather in Paris?<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "tool call with response",
+			msgs: []api.Message{
+				{Role: "user", Content: "What's the weather in Paris?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: map[string]any{"city": "Paris"},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Sunny, 72F"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get the current weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:     "object",
+							Required: []string{"city"},
+							Properties: map[string]api.ToolProperty{
+								"city": {Type: api.PropertyType{"string"}, Description: "The city name"},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n" +
+				"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
+				"<function>\n<name>get_weather</name>\n" +
+				"<description>Get the current weather</description>\n" +
+				"<parameters>\n" +
+				"<parameter>\n<name>city</name>\n<type>string</type>\n<description>The city name</description>\n</parameter>\n" +
+				"<required>[\"city\"]</required>\n" +
+				"</parameters>\n</function>\n</tools>\n\n" +
+				"If you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+				"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+				"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+				"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+				"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+				"- Required parameters MUST be specified\n" +
+				"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+				"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
+				"</IMPORTANT><|im_end|>\n" +
+				"<|im_start|>user\nWhat's the weather in Paris?<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
+				"<|im_start|>user\n<tool_response>\nSunny, 72F\n</tool_response>\n<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "assistant with content and tool call",
+			msgs: []api.Message{
+				{Role: "user", Content: "What's the weather?"},
+				{
+					Role:    "assistant",
+					Content: "Let me check that for you.",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: map[string]any{"city": "Paris"},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Sunny"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "get_weather",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"city": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n" +
+				"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
+				"<function>\n<name>get_weather</name>\n" +
+				"<parameters>\n" +
+				"<parameter>\n<name>city</name>\n<type>string</type>\n</parameter>\n" +
+				"</parameters>\n</function>\n</tools>\n\n" +
+				"If you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+				"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+				"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+				"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+				"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+				"- Required parameters MUST be specified\n" +
+				"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+				"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
+				"</IMPORTANT><|im_end|>\n" +
+				"<|im_start|>user\nWhat's the weather?<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>Let me check that for you.\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
+				"<|im_start|>user\n<tool_response>\nSunny\n</tool_response>\n<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "thinking in history is truncated",
+			msgs: []api.Message{
+				{Role: "user", Content: "Hi"},
+				{Role: "assistant", Content: "Hello!", Thinking: "Let me think about this..."},
+				{Role: "user", Content: "How are you?"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n<|im_end|>\n" +
+				"<|im_start|>user\nHi<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>Hello!<|im_end|>\n" +
+				"<|im_start|>user\nHow are you?<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "parallel tool calls",
+			msgs: []api.Message{
+				{Role: "user", Content: "Weather in Paris and London?"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: map[string]any{"city": "Paris"},
+							},
+						},
+						{
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: map[string]any{"city": "London"},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: "Sunny"},
+				{Role: "tool", Content: "Rainy"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "get_weather",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"city": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n" +
+				"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
+				"<function>\n<name>get_weather</name>\n" +
+				"<parameters>\n" +
+				"<parameter>\n<name>city</name>\n<type>string</type>\n</parameter>\n" +
+				"</parameters>\n</function>\n</tools>\n\n" +
+				"If you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+				"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+				"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+				"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+				"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+				"- Required parameters MUST be specified\n" +
+				"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+				"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
+				"</IMPORTANT><|im_end|>\n" +
+				"<|im_start|>user\nWeather in Paris and London?<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nLondon\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
+				"<|im_start|>user\n<tool_response>\nSunny\n</tool_response>\n<tool_response>\nRainy\n</tool_response>\n<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "thinking disabled when user doesn't request it",
+			msgs: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			thinkValue: nil,
+			expected: "<|im_start|>system\n<|im_end|>\n" +
+				"<|im_start|>user\nHello!<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>",
+		},
+		{
+			name: "complex message history with thinking, tools, tool calls, tool results and content",
+			msgs: []api.Message{
+				{Role: "user", Content: "What's the weather in Paris and London? Also, what's 2+2?"},
+				{Role: "assistant", Content: "", Thinking: "I need to check the weather for both cities and calculate 2+2. Let me start with the weather calls.", ToolCalls: []api.ToolCall{
+					{Function: api.ToolCallFunction{Name: "get_weather", Arguments: api.ToolCallFunctionArguments{"city": "Paris"}}},
+					{Function: api.ToolCallFunction{Name: "get_weather", Arguments: api.ToolCallFunctionArguments{"city": "London"}}},
+				}},
+				{Role: "tool", Content: "Sunny, 22°C", ToolCallID: "call1"},
+				{Role: "tool", Content: "Rainy, 15°C", ToolCallID: "call2"},
+				{Role: "assistant", Content: "", Thinking: "Now I have the weather data. Let me calculate 2+2.", ToolCalls: []api.ToolCall{
+					{Function: api.ToolCallFunction{Name: "calculate", Arguments: api.ToolCallFunctionArguments{"expression": "2+2"}}},
+				}},
+				{Role: "tool", Content: "4", ToolCallID: "call3"},
+				{Role: "assistant", Content: "Based on the weather data, Paris is sunny at 22°C and London is rainy at 15°C. Also, 2+2 equals 4.", Thinking: "Perfect! I have all the information needed to provide a complete answer."},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "get_weather",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"city": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "calculate",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"expression": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n" +
+				"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
+				"<function>\n<name>get_weather</name>\n" +
+				"<parameters>\n" +
+				"<parameter>\n<name>city</name>\n<type>string</type>\n</parameter>\n" +
+				"</parameters>\n</function>\n" +
+				"<function>\n<name>calculate</name>\n" +
+				"<parameters>\n" +
+				"<parameter>\n<name>expression</name>\n<type>string</type>\n</parameter>\n" +
+				"</parameters>\n</function>\n</tools>\n\n" +
+				"If you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+				"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+				"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+				"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+				"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+				"- Required parameters MUST be specified\n" +
+				"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+				"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
+				"</IMPORTANT><|im_end|>\n" +
+				"<|im_start|>user\nWhat's the weather in Paris and London? Also, what's 2+2?<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				"<think>\nI need to check the weather for both cities and calculate 2+2. Let me start with the weather calls.\n</think>\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nParis\n</parameter>\n</function>\n</tool_call>\n" +
+				"<tool_call>\n<function=get_weather>\n<parameter=city>\nLondon\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
+				"<|im_start|>user\n<tool_response>\nSunny, 22°C\n</tool_response>\n<tool_response>\nRainy, 15°C\n</tool_response>\n<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				"<think>\nNow I have the weather data. Let me calculate 2+2.\n</think>\n" +
+				"<tool_call>\n<function=calculate>\n<parameter=expression>\n2+2\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
+				"<|im_start|>user\n<tool_response>\n4\n</tool_response>\n<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				"<think>\nPerfect! I have all the information needed to provide a complete answer.\n</think>\n" +
+				"Based on the weather data, Paris is sunny at 22°C and London is rainy at 15°C. Also, 2+2 equals 4.<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name:       "empty messages list",
+			msgs:       []api.Message{},
+			thinkValue: nil,
+			expected:   "<|im_start|>system\n<|im_end|>\n<|im_start|>assistant\n<think></think>",
+		},
+		{
+			name: "tool result with JSON content",
+			msgs: []api.Message{
+				{Role: "user", Content: "Get user info"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{Function: api.ToolCallFunction{Name: "get_user", Arguments: map[string]any{"id": "123"}}},
+					},
+				},
+				{Role: "tool", Content: `{"name": "John", "age": 30, "active": true}`},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "get_user",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Properties: map[string]api.ToolProperty{"id": {Type: api.PropertyType{"string"}}},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n" +
+				"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
+				"<function>\n<name>get_user</name>\n<parameters>\n" +
+				"<parameter>\n<name>id</name>\n<type>string</type>\n</parameter>\n" +
+				"</parameters>\n</function>\n</tools>\n\n" +
+				"If you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+				"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+				"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+				"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+				"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+				"- Required parameters MUST be specified\n" +
+				"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+				"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
+				"</IMPORTANT><|im_end|>\n" +
+				"<|im_start|>user\nGet user info<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>\n" +
+				"<tool_call>\n<function=get_user>\n<parameter=id>\n123\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
+				"<|im_start|>user\n<tool_response>\n{\"name\": \"John\", \"age\": 30, \"active\": true}\n</tool_response>\n<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "assistant message with only thinking no content",
+			msgs: []api.Message{
+				{Role: "user", Content: "Think about this"},
+				{Role: "assistant", Thinking: "Deep thoughts here...", Content: ""},
+				{Role: "user", Content: "What did you think?"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n<|im_end|>\n" +
+				"<|im_start|>user\nThink about this<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think><|im_end|>\n" +
+				"<|im_start|>user\nWhat did you think?<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "tool call with complex nested argument",
+			msgs: []api.Message{
+				{Role: "user", Content: "Create data"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{Function: api.ToolCallFunction{
+							Name: "create",
+							Arguments: map[string]any{
+								"data": map[string]any{"nested": "value", "count": 42},
+							},
+						}},
+					},
+				},
+				{Role: "tool", Content: "Created"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "create",
+						Parameters: api.ToolFunctionParameters{
+							Type:       "object",
+							Properties: map[string]api.ToolProperty{"data": {Type: api.PropertyType{"object"}}},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n" +
+				"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
+				"<function>\n<name>create</name>\n<parameters>\n" +
+				"<parameter>\n<name>data</name>\n<type>object</type>\n</parameter>\n" +
+				"</parameters>\n</function>\n</tools>\n\n" +
+				"If you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+				"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+				"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+				"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+				"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+				"- Required parameters MUST be specified\n" +
+				"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+				"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
+				"</IMPORTANT><|im_end|>\n" +
+				"<|im_start|>user\nCreate data<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>\n" +
+				"<tool_call>\n<function=create>\n<parameter=data>\n{\"count\":42,\"nested\":\"value\"}\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
+				"<|im_start|>user\n<tool_response>\nCreated\n</tool_response>\n<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "content explaining the format itself",
+			msgs: []api.Message{
+				{Role: "user", Content: "How do I format a tool call?"},
+				{Role: "assistant", Content: "To call a tool, use <tool_call> tags with <function=name> inside."},
+				{Role: "user", Content: "Thanks!"},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n<|im_end|>\n" +
+				"<|im_start|>user\nHow do I format a tool call?<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>To call a tool, use <tool_call> tags with <function=name> inside.<|im_end|>\n" +
+				"<|im_start|>user\nThanks!<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+		{
+			name: "unicode in content and tool args",
+			msgs: []api.Message{
+				{Role: "user", Content: "Translate 你好"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{Function: api.ToolCallFunction{Name: "translate", Arguments: map[string]any{"text": "你好"}}},
+					},
+				},
+				{Role: "tool", Content: "Hello"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "translate",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"text": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+			thinkValue: &api.ThinkValue{Value: true},
+			expected: "<|im_start|>system\n" +
+				"# Tools\n\nYou have access to the following functions:\n\n<tools>\n" +
+				"<function>\n<name>translate</name>\n<parameters>\n" +
+				"<parameter>\n<name>text</name>\n<type>string</type>\n</parameter>\n" +
+				"</parameters>\n</function>\n</tools>\n\n" +
+				"If you choose to call a function ONLY reply in the following format with NO suffix:\n\n" +
+				"<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n" +
+				"<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n" +
+				"</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n" +
+				"- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n" +
+				"- Required parameters MUST be specified\n" +
+				"- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n" +
+				"- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n" +
+				"</IMPORTANT><|im_end|>\n" +
+				"<|im_start|>user\nTranslate 你好<|im_end|>\n" +
+				"<|im_start|>assistant\n<think></think>\n" +
+				"<tool_call>\n<function=translate>\n<parameter=text>\n你好\n</parameter>\n</function>\n</tool_call>\n<|im_end|>\n" +
+				"<|im_start|>user\n<tool_response>\nHello\n</tool_response>\n<|im_end|>\n" +
+				"<|im_start|>assistant\n<think>\n",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			renderer := &Nemotron3NanoRenderer{}
+			rendered, err := renderer.Render(tt.msgs, tt.tools, tt.thinkValue)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if diff := cmp.Diff(rendered, tt.expected); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}

model/renderers/olmo3.go

@@ -10,12 +10,15 @@ import (
 )
 
 const (
-	olmo3DefaultSystemMessage = "You are a helpful function-calling AI assistant. "
-	olmo3NoFunctionsMessage   = "You do not currently have access to any functions. "
-	olmo3WithFunctionsMessage = "You are provided with function signatures within <functions></functions> XML tags. You may call one or more functions to assist with the user query. Output any function calls within <function_calls></function_calls> XML tags. Do not make assumptions about what values to plug into functions."
+	olmo3DefaultSystemMessage  = "You are a helpful function-calling AI assistant. "
+	olmo31DefaultSystemMessage = "You are Olmo, a helpful AI assistant built by Ai2. Your date cutoff is December 2024, and your model weights are available at https://huggingface.co/allenai. "
+	olmo3NoFunctionsMessage    = "You do not currently have access to any functions. "
+	olmo3WithFunctionsMessage  = "You are provided with function signatures within <functions></functions> XML tags. You may call one or more functions to assist with the user query. Output any function calls within <function_calls></function_calls> XML tags. Do not make assumptions about what values to plug into functions."
 )
 
-type Olmo3Renderer struct{}
+type Olmo3Renderer struct {
+	UseExtendedSystemMessage bool
+}
 
 func (r *Olmo3Renderer) Render(messages []api.Message, tools []api.Tool, _ *api.ThinkValue) (string, error) {
 	var sb strings.Builder
@@ -51,7 +54,11 @@ func (r *Olmo3Renderer) Render(messages []api.Message, tools []api.Tool, _ *api.
 	} else {
 		// Default system message - single newline after "system"
 		sb.WriteString("<|im_start|>system\n")
-		sb.WriteString(olmo3DefaultSystemMessage)
+		if r.UseExtendedSystemMessage {
+			sb.WriteString(olmo31DefaultSystemMessage)
+		} else {
+			sb.WriteString(olmo3DefaultSystemMessage)
+		}
 
 		if len(tools) > 0 {
 			functionsJSON, err := marshalWithSpaces(tools)
@@ -140,7 +147,7 @@ func (r *Olmo3Renderer) Render(messages []api.Message, tools []api.Tool, _ *api.
 	}
 
 	if needsGenerationPrompt {
-		sb.WriteString("<|im_start|>assistant\n\n")
+		sb.WriteString("<|im_start|>assistant\n")
 	}
 
 	return sb.String(), nil

model/renderers/olmo3_test.go

@@ -24,7 +24,7 @@ func TestOlmo3Renderer(t *testing.T) {
 				"You are a helpful function-calling AI assistant. You do not currently have access to any functions. <functions></functions><|im_end|>\n" +
 				"<|im_start|>user\n" +
 				"Hello!<|im_end|>\n" +
-				"<|im_start|>assistant\n\n",
+				"<|im_start|>assistant\n",
 		},
 		{
 			name: "with system message no tools",
@@ -36,7 +36,7 @@ func TestOlmo3Renderer(t *testing.T) {
 				"You are a helpful assistant.<|im_end|>\n" +
 				"<|im_start|>user\n" +
 				"Hello!<|im_end|>\n" +
-				"<|im_start|>assistant\n\n",
+				"<|im_start|>assistant\n",
 		},
 		{
 			name: "with system message and tools",
@@ -64,7 +64,7 @@ func TestOlmo3Renderer(t *testing.T) {
 				`You are a helpful assistant.<functions>[{"type": "function", "function": {"name": "get_weather", "description": "Get the current weather", "parameters": {"type": "object", "required": ["location"], "properties": {"location": {"type": "string", "description": "The city"}}}}}]</functions><|im_end|>` + "\n" +
 				"<|im_start|>user\n" +
 				"What is the weather?<|im_end|>\n" +
-				"<|im_start|>assistant\n\n",
+				"<|im_start|>assistant\n",
 		},
 		{
 			name: "default system with tools - includes function instruction",
@@ -93,7 +93,7 @@ func TestOlmo3Renderer(t *testing.T) {
 				`<functions>[{"type": "function", "function": {"name": "get_weather", "description": "Get the current weather", "parameters": {"type": "object", "required": ["location"], "properties": {"location": {"type": "string", "description": "The city"}}}}}]</functions><|im_end|>` + "\n" +
 				"<|im_start|>user\n" +
 				"What is the weather?<|im_end|>\n" +
-				"<|im_start|>assistant\n\n",
+				"<|im_start|>assistant\n",
 		},
 		{
 			name: "assistant with tool calls - function call syntax",
@@ -141,7 +141,7 @@ func TestOlmo3Renderer(t *testing.T) {
 				`Let me check the weather.<function_calls>get_weather(location="San Francisco")</function_calls><|im_end|>` + "\n" +
 				"<|im_start|>environment\n" +
 				`{"temperature": 68}<|im_end|>` + "\n" +
-				"<|im_start|>assistant\n\n",
+				"<|im_start|>assistant\n",
 		},
 		{
 			name: "multi-turn conversation",
@@ -159,7 +159,7 @@ func TestOlmo3Renderer(t *testing.T) {
 				"Hi there!<|im_end|>\n" +
 				"<|im_start|>user\n" +
 				"How are you?<|im_end|>\n" +
-				"<|im_start|>assistant\n\n",
+				"<|im_start|>assistant\n",
 		},
 		{
 			name: "parallel tool calls - newline separated",
@@ -214,7 +214,7 @@ func TestOlmo3Renderer(t *testing.T) {
 				`{"temperature": 68}<|im_end|>` + "\n" +
 				"<|im_start|>environment\n" +
 				`{"temperature": 55}<|im_end|>` + "\n" +
-				"<|im_start|>assistant\n\n",
+				"<|im_start|>assistant\n",
 		},
 		{
 			name: "tool call with multiple arguments",
@@ -259,7 +259,7 @@ func TestOlmo3Renderer(t *testing.T) {
 				"Book a flight<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
 				`<function_calls>book_flight(from="SFO", to="NYC")</function_calls><|im_end|>` + "\n" +
-				"<|im_start|>assistant\n\n",
+				"<|im_start|>assistant\n",
 		},
 		{
 			name: "assistant prefill - no generation prompt",

model/renderers/olmo3_think.go

@@ -1,31 +1,31 @@
 package renderers
 
 import (
-	"encoding/json"
 	"strings"
 
 	"github.com/ollama/ollama/api"
 )
 
+type Olmo3ThinkVariant int
+
 const (
-	olmo3ThinkDefaultSystemMessage = "You are OLMo, a helpful function-calling AI assistant built by Ai2. Your date cutoff is November 2024, and your model weights are available at https://huggingface.co/allenai."
-	olmo3ThinkNoFunctionsMessage   = " You do not currently have access to any functions."
+	// Olmo3Think32B is for allenai/Olmo-3-32B-Think
+	Olmo3Think32B Olmo3ThinkVariant = iota
+	// Olmo31Think is for allenai/Olmo-3-7B-Think and allenai/Olmo-3.1-32B-Think (includes model info)
+	Olmo31Think
 )
 
-type Olmo3ThinkRenderer struct{}
+const (
+	olmo3ThinkFunctionsSuffix  = " You do not currently have access to any functions. <functions></functions>"
+	olmo3Think32BSystemMessage = "You are a helpful AI assistant."
+	olmo31ThinkSystemMessage   = "You are Olmo, a helpful AI assistant built by Ai2. Your date cutoff is December 2024, and your model weights are available at https://huggingface.co/allenai."
+)
 
-type olmo3ThinkToolCall struct {
-	ID       string                 `json:"id,omitempty"`
-	Type     string                 `json:"type,omitempty"`
-	Function olmo3ThinkToolCallFunc `json:"function"`
+type Olmo3ThinkRenderer struct {
+	Variant Olmo3ThinkVariant
 }
 
-type olmo3ThinkToolCallFunc struct {
-	Name      string `json:"name"`
-	Arguments string `json:"arguments"`
-}
-
-func (r *Olmo3ThinkRenderer) Render(messages []api.Message, tools []api.Tool, _ *api.ThinkValue) (string, error) {
+func (r *Olmo3ThinkRenderer) Render(messages []api.Message, _ []api.Tool, _ *api.ThinkValue) (string, error) {
 	var sb strings.Builder
 
 	var systemMessage *api.Message
@@ -37,34 +37,31 @@ func (r *Olmo3ThinkRenderer) Render(messages []api.Message, tools []api.Tool, _
 			}
 			continue
 		}
+		// Skip tool messages - Think models don't support tools
+		if message.Role == "tool" {
+			continue
+		}
 		filteredMessages = append(filteredMessages, message)
 	}
 
-	systemContent := olmo3ThinkDefaultSystemMessage
-	if systemMessage != nil {
-		systemContent = systemMessage.Content
-	}
-
 	sb.WriteString("<|im_start|>system\n")
-	sb.WriteString(systemContent)
 
-	if len(tools) > 0 {
-		functionsJSON, err := marshalWithSpaces(tools)
-		if err != nil {
-			return "", err
-		}
-		sb.WriteString(" <functions>")
-		sb.WriteString(string(functionsJSON))
-		sb.WriteString("</functions>")
+	if systemMessage != nil {
+		sb.WriteString(systemMessage.Content)
+		sb.WriteString(olmo3ThinkFunctionsSuffix)
 	} else {
-		sb.WriteString(olmo3ThinkNoFunctionsMessage)
-		sb.WriteString(" <functions></functions>")
+		// Default system message varies by variant
+		switch r.Variant {
+		case Olmo3Think32B:
+			sb.WriteString(olmo3Think32BSystemMessage)
+		default: // Olmo3Think7B, Olmo31Think use same template - diverges from HF but confirmed difference from team
+			sb.WriteString(olmo31ThinkSystemMessage)
+		}
 	}
+
 	sb.WriteString("<|im_end|>\n")
 
-	for i, message := range filteredMessages {
-		lastMessage := i == len(filteredMessages)-1
-
+	for _, message := range filteredMessages {
 		switch message.Role {
 		case "user":
 			sb.WriteString("<|im_start|>user\n")
@@ -73,58 +70,15 @@ func (r *Olmo3ThinkRenderer) Render(messages []api.Message, tools []api.Tool, _
 
 		case "assistant":
 			sb.WriteString("<|im_start|>assistant\n")
-
 			if message.Content != "" {
 				sb.WriteString(message.Content)
 			}
-
-			if len(message.ToolCalls) > 0 {
-				toolCalls := make([]olmo3ThinkToolCall, len(message.ToolCalls))
-				for j, tc := range message.ToolCalls {
-					argsJSON, err := json.Marshal(tc.Function.Arguments)
-					if err != nil {
-						return "", err
-					}
-					toolCalls[j] = olmo3ThinkToolCall{
-						ID:   tc.ID,
-						Type: "function",
-						Function: olmo3ThinkToolCallFunc{
-							Name:      tc.Function.Name,
-							Arguments: string(argsJSON),
-						},
-					}
-				}
-				toolCallsJSON, err := marshalWithSpaces(toolCalls)
-				if err != nil {
-					return "", err
-				}
-				sb.WriteString("<function_calls>")
-				sb.WriteString(string(toolCallsJSON))
-				sb.WriteString("</function_calls>")
-			}
-
-			if !lastMessage {
-				sb.WriteString("<|im_end|>\n")
-			}
-
-		case "tool":
-			sb.WriteString("<|im_start|>environment\n")
-			sb.WriteString(message.Content)
 			sb.WriteString("<|im_end|>\n")
 		}
 	}
 
-	needsGenerationPrompt := true
-	if len(filteredMessages) > 0 {
-		lastMsg := filteredMessages[len(filteredMessages)-1]
-		if lastMsg.Role == "assistant" && len(lastMsg.ToolCalls) == 0 && lastMsg.Content != "" {
-			needsGenerationPrompt = false
-		}
-	}
-
-	if needsGenerationPrompt {
-		sb.WriteString("<|im_start|>assistant\n<think>")
-	}
+	// Always add generation prompt with <think> tag for thinking models
+	sb.WriteString("<|im_start|>assistant\n<think>")
 
 	return sb.String(), nil
 }

model/renderers/olmo3_think_test.go

@@ -11,24 +11,27 @@ import (
 func TestOlmo3ThinkRenderer(t *testing.T) {
 	tests := []struct {
 		name     string
+		variant  Olmo3ThinkVariant
 		msgs     []api.Message
 		tools    []api.Tool
 		expected string
 	}{
 		{
-			name: "basic without system - adds default system",
+			name:    "7b_basic_without_system",
+			variant: Olmo31Think,
 			msgs: []api.Message{
 				{Role: "user", Content: "Hello!"},
 			},
 			expected: "<|im_start|>system\n" +
-				"You are OLMo, a helpful function-calling AI assistant built by Ai2. Your date cutoff is November 2024, and your model weights are available at https://huggingface.co/allenai. You do not currently have access to any functions. <functions></functions><|im_end|>\n" +
+				"You are Olmo, a helpful AI assistant built by Ai2. Your date cutoff is December 2024, and your model weights are available at https://huggingface.co/allenai.<|im_end|>\n" +
 				"<|im_start|>user\n" +
 				"Hello!<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
 				"<think>",
 		},
 		{
-			name: "with system message no tools",
+			name:    "7b_with_custom_system",
+			variant: Olmo31Think,
 			msgs: []api.Message{
 				{Role: "system", Content: "You are a helpful assistant."},
 				{Role: "user", Content: "Hello!"},
@@ -41,9 +44,9 @@ func TestOlmo3ThinkRenderer(t *testing.T) {
 				"<think>",
 		},
 		{
-			name: "with system message and tools",
+			name:    "7b_tools_ignored",
+			variant: Olmo31Think,
 			msgs: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant."},
 				{Role: "user", Content: "What is the weather?"},
 			},
 			tools: []api.Tool{
@@ -52,27 +55,20 @@ func TestOlmo3ThinkRenderer(t *testing.T) {
 					Function: api.ToolFunction{
 						Name:        "get_weather",
 						Description: "Get the current weather",
-						Parameters: api.ToolFunctionParameters{
-							Type:     "object",
-							Required: []string{"location"},
-							Properties: map[string]api.ToolProperty{
-								"location": {Type: api.PropertyType{"string"}, Description: "The city"},
-							},
-						},
 					},
 				},
 			},
 			expected: "<|im_start|>system\n" +
-				`You are a helpful assistant. <functions>[{"type": "function", "function": {"name": "get_weather", "description": "Get the current weather", "parameters": {"type": "object", "required": ["location"], "properties": {"location": {"type": "string", "description": "The city"}}}}}]</functions><|im_end|>` + "\n" +
+				"You are Olmo, a helpful AI assistant built by Ai2. Your date cutoff is December 2024, and your model weights are available at https://huggingface.co/allenai.<|im_end|>\n" +
 				"<|im_start|>user\n" +
 				"What is the weather?<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
 				"<think>",
 		},
 		{
-			name: "assistant with tool calls",
+			name:    "7b_tool_calls_and_tool_messages_ignored",
+			variant: Olmo31Think,
 			msgs: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant."},
 				{Role: "user", Content: "What is the weather in SF?"},
 				{
 					Role:    "assistant",
@@ -81,53 +77,33 @@ func TestOlmo3ThinkRenderer(t *testing.T) {
 						{
 							ID: "call_1",
 							Function: api.ToolCallFunction{
-								Name: "get_weather",
-								Arguments: map[string]any{
-									"location": "San Francisco",
-								},
-							},
-						},
-					},
-				},
-				{Role: "tool", Content: `{"temperature": 68}`, ToolName: "get_weather"},
-			},
-			tools: []api.Tool{
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name:        "get_weather",
-						Description: "Get the current weather",
-						Parameters: api.ToolFunctionParameters{
-							Type:     "object",
-							Required: []string{"location"},
-							Properties: map[string]api.ToolProperty{
-								"location": {Type: api.PropertyType{"string"}, Description: "The city"},
+								Name:      "get_weather",
+								Arguments: map[string]any{"location": "San Francisco"},
 							},
 						},
 					},
 				},
+				{Role: "tool", Content: `{"temperature": 68}`},
 			},
 			expected: "<|im_start|>system\n" +
-				`You are a helpful assistant. <functions>[{"type": "function", "function": {"name": "get_weather", "description": "Get the current weather", "parameters": {"type": "object", "required": ["location"], "properties": {"location": {"type": "string", "description": "The city"}}}}}]</functions><|im_end|>` + "\n" +
+				"You are Olmo, a helpful AI assistant built by Ai2. Your date cutoff is December 2024, and your model weights are available at https://huggingface.co/allenai.<|im_end|>\n" +
 				"<|im_start|>user\n" +
 				"What is the weather in SF?<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
-				`Let me check the weather.<function_calls>[{"id": "call_1", "type": "function", "function": {"name": "get_weather", "arguments": "{\"location\":\"San Francisco\"}"}}]</function_calls><|im_end|>` + "\n" +
-				"<|im_start|>environment\n" +
-				`{"temperature": 68}<|im_end|>` + "\n" +
+				"Let me check the weather.<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
 				"<think>",
 		},
 		{
-			name: "multi-turn conversation",
+			name:    "7b_multi_turn_conversation",
+			variant: Olmo31Think,
 			msgs: []api.Message{
-				{Role: "system", Content: "You are a helpful assistant."},
 				{Role: "user", Content: "Hello"},
 				{Role: "assistant", Content: "Hi there!"},
 				{Role: "user", Content: "How are you?"},
 			},
 			expected: "<|im_start|>system\n" +
-				"You are a helpful assistant. You do not currently have access to any functions. <functions></functions><|im_end|>\n" +
+				"You are Olmo, a helpful AI assistant built by Ai2. Your date cutoff is December 2024, and your model weights are available at https://huggingface.co/allenai.<|im_end|>\n" +
 				"<|im_start|>user\n" +
 				"Hello<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
@@ -138,73 +114,56 @@ func TestOlmo3ThinkRenderer(t *testing.T) {
 				"<think>",
 		},
 		{
-			name: "parallel tool calls",
+			name:    "32b_basic_without_system",
+			variant: Olmo3Think32B,
 			msgs: []api.Message{
-				{Role: "user", Content: "Get weather in SF and NYC"},
-				{
-					Role: "assistant",
-					ToolCalls: []api.ToolCall{
-						{
-							ID: "call_1",
-							Function: api.ToolCallFunction{
-								Name:      "get_weather",
-								Arguments: map[string]any{"location": "San Francisco"},
-							},
-						},
-						{
-							ID: "call_2",
-							Function: api.ToolCallFunction{
-								Name:      "get_weather",
-								Arguments: map[string]any{"location": "New York"},
-							},
-						},
-					},
-				},
-				{Role: "tool", Content: `{"temperature": 68}`, ToolName: "get_weather"},
-				{Role: "tool", Content: `{"temperature": 55}`, ToolName: "get_weather"},
-			},
-			tools: []api.Tool{
-				{
-					Type: "function",
-					Function: api.ToolFunction{
-						Name: "get_weather",
-						Parameters: api.ToolFunctionParameters{
-							Type: "object",
-							Properties: map[string]api.ToolProperty{
-								"location": {Type: api.PropertyType{"string"}},
-							},
-						},
-					},
-				},
+				{Role: "user", Content: "Hello!"},
 			},
 			expected: "<|im_start|>system\n" +
-				`You are OLMo, a helpful function-calling AI assistant built by Ai2. Your date cutoff is November 2024, and your model weights are available at https://huggingface.co/allenai. <functions>[{"type": "function", "function": {"name": "get_weather", "parameters": {"type": "object", "properties": {"location": {"type": "string"}}}}}]</functions><|im_end|>` + "\n" +
+				"You are a helpful AI assistant.<|im_end|>\n" +
 				"<|im_start|>user\n" +
-				"Get weather in SF and NYC<|im_end|>\n" +
-				"<|im_start|>assistant\n" +
-				`<function_calls>[{"id": "call_1", "type": "function", "function": {"name": "get_weather", "arguments": "{\"location\":\"San Francisco\"}"}}, {"id": "call_2", "type": "function", "function": {"name": "get_weather", "arguments": "{\"location\":\"New York\"}"}}]</function_calls><|im_end|>` + "\n" +
-				"<|im_start|>environment\n" +
-				`{"temperature": 68}<|im_end|>` + "\n" +
-				"<|im_start|>environment\n" +
-				`{"temperature": 55}<|im_end|>` + "\n" +
+				"Hello!<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
 				"<think>",
 		},
 		{
-			name: "assistant message only content no tool calls",
+			name:    "32b_with_custom_system_gets_suffix",
+			variant: Olmo3Think32B,
 			msgs: []api.Message{
-				{Role: "user", Content: "Tell me a joke"},
-				{Role: "assistant", Content: "Why did the chicken cross the road?"},
-				{Role: "user", Content: "I don't know, why?"},
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello!"},
 			},
 			expected: "<|im_start|>system\n" +
-				"You are OLMo, a helpful function-calling AI assistant built by Ai2. Your date cutoff is November 2024, and your model weights are available at https://huggingface.co/allenai. You do not currently have access to any functions. <functions></functions><|im_end|>\n" +
+				"You are a helpful assistant. You do not currently have access to any functions. <functions></functions><|im_end|>\n" +
 				"<|im_start|>user\n" +
-				"Tell me a joke<|im_end|>\n" +
+				"Hello!<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
-				"Why did the chicken cross the road?<|im_end|>\n" +
+				"<think>",
+		},
+		{
+			name:    "31_basic_without_system",
+			variant: Olmo31Think,
+			msgs: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are Olmo, a helpful AI assistant built by Ai2. Your date cutoff is December 2024, and your model weights are available at https://huggingface.co/allenai.<|im_end|>\n" +
 				"<|im_start|>user\n" +
-				"I don't know, why?<|im_end|>\n" +
+				"Hello!<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				"<think>",
+		},
+		{
+			name:    "31_with_custom_system_gets_suffix",
+			variant: Olmo31Think,
+			msgs: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello!"},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are a helpful assistant. You do not currently have access to any functions. <functions></functions><|im_end|>\n" +
+				"<|im_start|>user\n" +
+				"Hello!<|im_end|>\n" +
 				"<|im_start|>assistant\n" +
 				"<think>",
 		},
@@ -212,7 +171,7 @@ func TestOlmo3ThinkRenderer(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			rendered, err := (&Olmo3ThinkRenderer{}).Render(tt.msgs, tt.tools, nil)
+			rendered, err := (&Olmo3ThinkRenderer{Variant: tt.variant}).Render(tt.msgs, tt.tools, nil)
 			if err != nil {
 				t.Fatal(err)
 			}

model/renderers/renderer.go

@@ -59,12 +59,25 @@ func rendererForName(name string) Renderer {
 	case "cogito":
 		renderer := &CogitoRenderer{isThinking: true}
 		return renderer
+	case "deepseek-v3.1":
+		renderer := &DeepSeek3Renderer{IsThinking: true, Variant: Deepseek31}
+		return renderer
 	case "olmo3":
-		renderer := &Olmo3Renderer{}
+		renderer := &Olmo3Renderer{UseExtendedSystemMessage: false}
+		return renderer
+	case "olmo3.1":
+		renderer := &Olmo3Renderer{UseExtendedSystemMessage: true}
 		return renderer
 	case "olmo3-think":
-		renderer := &Olmo3ThinkRenderer{}
+		// Used for Olmo-3-7B-Think and Olmo-3.1-32B-Think (same template)
+		renderer := &Olmo3ThinkRenderer{Variant: Olmo31Think}
 		return renderer
+	case "olmo3-32b-think":
+		// Used for Olmo-3-32B-Think
+		renderer := &Olmo3ThinkRenderer{Variant: Olmo3Think32B}
+		return renderer
+	case "nemotron-3-nano":
+		return &Nemotron3NanoRenderer{}
 	default:
 		return nil
 	}

openai/openai.go

@@ -487,29 +487,9 @@ func FromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 						}
 					}
 
-					types := []string{"jpeg", "jpg", "png", "webp"}
-					valid := false
-					// support blank mime type to match api/chat taking just unadorned base64
-					if strings.HasPrefix(url, "data:;base64,") {
-						url = strings.TrimPrefix(url, "data:;base64,")
-						valid = true
-					}
-					for _, t := range types {
-						prefix := "data:image/" + t + ";base64,"
-						if strings.HasPrefix(url, prefix) {
-							url = strings.TrimPrefix(url, prefix)
-							valid = true
-							break
-						}
-					}
-
-					if !valid {
-						return nil, errors.New("invalid image input")
-					}
-
-					img, err := base64.StdEncoding.DecodeString(url)
+					img, err := decodeImageURL(url)
 					if err != nil {
-						return nil, errors.New("invalid message format")
+						return nil, err
 					}
 
 					messages = append(messages, api.Message{Role: msg.Role, Images: []api.ImageData{img}})
@@ -648,6 +628,35 @@ func nameFromToolCallID(messages []Message, toolCallID string) string {
 	return ""
 }
 
+// decodeImageURL decodes a base64 data URI into raw image bytes.
+func decodeImageURL(url string) (api.ImageData, error) {
+	types := []string{"jpeg", "jpg", "png", "webp"}
+
+	// Support blank mime type to match /api/chat's behavior of taking just unadorned base64
+	if strings.HasPrefix(url, "data:;base64,") {
+		url = strings.TrimPrefix(url, "data:;base64,")
+	} else {
+		valid := false
+		for _, t := range types {
+			prefix := "data:image/" + t + ";base64,"
+			if strings.HasPrefix(url, prefix) {
+				url = strings.TrimPrefix(url, prefix)
+				valid = true
+				break
+			}
+		}
+		if !valid {
+			return nil, errors.New("invalid image input")
+		}
+	}
+
+	img, err := base64.StdEncoding.DecodeString(url)
+	if err != nil {
+		return nil, errors.New("invalid image input")
+	}
+	return img, nil
+}
+
 // FromCompletionToolCall converts OpenAI ToolCall format to api.ToolCall
 func FromCompletionToolCall(toolCalls []ToolCall) ([]api.ToolCall, error) {
 	apiToolCalls := make([]api.ToolCall, len(toolCalls))

runner/llamarunner/runner.go

@@ -26,6 +26,7 @@ import (
 	"github.com/ollama/ollama/llama"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
+	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/runner/common"
 )
 
@@ -832,7 +833,7 @@ func (s *Server) loadModel(
 	ppath string,
 	kvSize int,
 	kvCacheType string,
-	flashAttention bool,
+	flashAttention ml.FlashAttentionType,
 	threads int,
 	multiUserCache bool,
 ) {
@@ -842,7 +843,7 @@ func (s *Server) loadModel(
 		panic(err)
 	}
 
-	ctxParams := llama.NewContextParams(kvSize, s.batchSize*s.parallel, s.parallel, threads, flashAttention, kvCacheType)
+	ctxParams := llama.NewContextParams(kvSize, s.batchSize, s.parallel, threads, flashAttention, kvCacheType)
 	s.lc, err = llama.NewContextWithModel(s.model, ctxParams)
 	if err != nil {
 		panic(err)

runner/ollamarunner/runner.go

@@ -1203,16 +1203,22 @@ func (s *Server) allocModel(
 		return errors.New("loras are not yet implemented")
 	}
 
+	if s.model.Config().Cache == nil {
+		if parallel > 1 {
+			parallel = 1
+			slog.Warn("model does not support caching, disabling parallel processing")
+		}
+		if s.batchSize < kvSize {
+			s.batchSize = kvSize
+			slog.Warn("model does not support caching, setting batch size to context length", "batch_size", kvSize)
+		}
+	}
+
 	s.cache, err = NewInputCache(s.model, kvCacheType, int32(kvSize), parallel, s.batchSize, multiUserCache)
 	if err != nil {
 		return err
 	}
 
-	if !s.cache.enabled && parallel > 1 {
-		parallel = 1
-		slog.Warn("model does not support caching, disabling parallel processing")
-	}
-
 	s.parallel = parallel
 	s.seqs = make([]*Sequence, s.parallel)
 	s.seqsSem = semaphore.NewWeighted(int64(s.parallel))

server/routes.go

@@ -1532,6 +1532,7 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.POST("/v1/embeddings", middleware.EmbeddingsMiddleware(), s.EmbedHandler)
 	r.GET("/v1/models", middleware.ListMiddleware(), s.ListHandler)
 	r.GET("/v1/models/:model", middleware.RetrieveMiddleware(), s.ShowHandler)
+	r.POST("/v1/responses", middleware.ResponsesMiddleware(), s.ChatHandler)
 
 	if rc != nil {
 		// wrap old with new
@@ -2393,3 +2394,4 @@ func filterThinkTags(msgs []api.Message, m *Model) []api.Message {
 	}
 	return msgs
 }
+

template/template.go

@@ -127,6 +127,9 @@ var funcs = template.FuncMap{
 		// Default format is YYYY-MM-DD
 		return time.Now().Format("2006-01-02")
 	},
+	"yesterdayDate": func(args ...string) string {
+		return time.Now().AddDate(0, 0, -1).Format("2006-01-02")
+	},
 	"toTypeScriptType": func(v any) string {
 		if param, ok := v.(api.ToolProperty); ok {
 			return param.ToTypeScriptType()

template/template_test.go

@@ -10,6 +10,7 @@ import (
 	"slices"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/google/go-cmp/cmp"
 
@@ -451,6 +452,72 @@ func TestExecuteWithSuffix(t *testing.T) {
 	}
 }
 
+func TestDateFunctions(t *testing.T) {
+	t.Run("currentDate", func(t *testing.T) {
+		tmpl, err := Parse("{{- range .Messages }}{{ .Content }}{{ end }} Today is {{ currentDate }}")
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		var b bytes.Buffer
+		if err := tmpl.Execute(&b, Values{Messages: []api.Message{{Role: "user", Content: "Hello"}}}); err != nil {
+			t.Fatal(err)
+		}
+
+		expected := "Hello Today is " + time.Now().Format("2006-01-02")
+		if b.String() != expected {
+			t.Errorf("got %q, want %q", b.String(), expected)
+		}
+	})
+
+	t.Run("yesterdayDate", func(t *testing.T) {
+		tmpl, err := Parse("{{- range .Messages }}{{ .Content }}{{ end }} Yesterday was {{ yesterdayDate }}")
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		var b bytes.Buffer
+		if err := tmpl.Execute(&b, Values{Messages: []api.Message{{Role: "user", Content: "Hello"}}}); err != nil {
+			t.Fatal(err)
+		}
+
+		expected := "Hello Yesterday was " + time.Now().AddDate(0, 0, -1).Format("2006-01-02")
+		if b.String() != expected {
+			t.Errorf("got %q, want %q", b.String(), expected)
+		}
+	})
+
+	t.Run("yesterdayDate format", func(t *testing.T) {
+		tmpl, err := Parse("{{- range .Messages }}{{ end }}{{ yesterdayDate }}")
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		var b bytes.Buffer
+		if err := tmpl.Execute(&b, Values{Messages: []api.Message{{Role: "user", Content: "Hello"}}}); err != nil {
+			t.Fatal(err)
+		}
+
+		// Verify the format matches YYYY-MM-DD
+		result := b.String()
+		if len(result) != 10 {
+			t.Errorf("expected date length 10, got %d: %q", len(result), result)
+		}
+
+		// Parse and verify it's a valid date
+		parsed, err := time.Parse("2006-01-02", result)
+		if err != nil {
+			t.Errorf("failed to parse date %q: %v", result, err)
+		}
+
+		// Verify it's yesterday
+		yesterday := time.Now().AddDate(0, 0, -1)
+		if parsed.Year() != yesterday.Year() || parsed.Month() != yesterday.Month() || parsed.Day() != yesterday.Day() {
+			t.Errorf("expected yesterday's date, got %v", parsed)
+		}
+	})
+}
+
 func TestCollate(t *testing.T) {
 	cases := []struct {
 		name     string