remove cherry pick manually

This reverts commit 56f754f46b.

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

envconfig/config.go

@@ -199,7 +199,7 @@ var (
 	// MultiUserCache optimizes prompt caching for multi-user scenarios
 	MultiUserCache = Bool("OLLAMA_MULTIUSER_CACHE")
 	// Enable the new Ollama engine
-	NewEngine = BoolWithDefault("OLLAMA_NEW_ENGINE")
+	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
 	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
 	// Auth enables authentication between the Ollama client and server
@@ -291,7 +291,7 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
 		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
-		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(true), "Enable the new Ollama engine"},
+		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 		"OLLAMA_REMOTES":           {"OLLAMA_REMOTES", Remotes(), "Allowed hosts for remote models (default \"ollama.com\")"},
 
 		// Informational

llm/server.go

@@ -143,7 +143,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 	var llamaModel *llama.Model
 	var textProcessor model.TextProcessor
 	var err error
-	if envconfig.NewEngine(true) || f.KV().OllamaEngineRequired() {
+	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
 		if len(projectors) == 0 {
 			textProcessor, err = model.NewTextProcessor(modelPath)
 		} else {

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"
@@ -13,19 +12,20 @@ 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, scale float32) ml.Tensor {
@@ -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)),
@@ -77,28 +81,32 @@ 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")),
 			ropeExtrapolation:    c.Float("rope.scaling.extrapolation_factor", 1.0),
 			ropeBetaFast:         c.Float("rope.scaling.beta_fast", 64.0),
 			ropeBetaSlow:         c.Float("rope.scaling.beta_slow", 1.0),
-			ropeScale:            c.Float("rope.scaling.factor", 8.0),
+			ropeScale:            c.Float("rope.scaling.factor", 1.0),
 			finalLogitSoftcap:    c.Float("final_logit_softcapping", 0.0),
 		},
 	}
 
-	// Google's Gemma 3 release with sliding window attention does
-	// not use final logit softcapping, and so force it to 0.0
-	// The QAT weights for Gemma 3 also included an incorrect
-	// value for the rope scale, so we need to set it to 1.0 here.
-	// 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
-		m.TextConfig.ropeScale = 1.0
+	// 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 {

model/parsers/deepseek.go

@@ -0,0 +1,292 @@
+package parsers
+
+import (
+	"encoding/json"
+	"errors"
+	"log/slog"
+	"strings"
+	"unicode"
+
+	"github.com/ollama/ollama/api"
+)
+
+type DeepSeekParserState int
+
+const (
+	DeepSeekCollectingThinking DeepSeekParserState = iota
+	DeepSeekCollectingContent
+	DeepSeekCollectingToolCalls
+	DeepSeekCollectingToolOutput
+)
+
+const (
+	deepseekThinkingCloseTag   = "</think>"
+	deepseekToolCallsBeginTag  = "<｜tool▁calls▁begin｜>"
+	deepseekToolCallsEndTag    = "<｜tool▁calls▁end｜>"
+	deepseekToolCallBeginTag   = "<｜tool▁call▁begin｜>"
+	deepseekToolCallEndTag     = "<｜tool▁call▁end｜>"
+	deepseekToolSepTag         = "<｜tool▁sep｜>"
+	deepseekToolOutputBeginTag = "<｜tool▁output▁begin｜>"
+	deepseekToolOutputEndTag   = "<｜tool▁output▁end｜>"
+)
+
+type DeepSeekParser struct {
+	state              DeepSeekParserState
+	buffer             strings.Builder
+	hasThinkingSupport bool
+}
+
+func (p *DeepSeekParser) HasToolSupport() bool {
+	return true
+}
+
+func (p *DeepSeekParser) HasThinkingSupport() bool {
+	return p.hasThinkingSupport
+}
+
+func (p *DeepSeekParser) setInitialState(lastMessage *api.Message, tools []api.Tool, thinkValue *api.ThinkValue) {
+	prefill := lastMessage != nil && lastMessage.Role == "assistant"
+
+	// Check both model capability AND request preference
+	thinkingEnabled := p.HasThinkingSupport() && (thinkValue == nil || thinkValue.Bool())
+
+	if !thinkingEnabled {
+		p.state = DeepSeekCollectingContent
+		return
+	}
+
+	if prefill && lastMessage.Content != "" {
+		p.state = DeepSeekCollectingContent
+		return
+	}
+
+	p.state = DeepSeekCollectingThinking
+}
+
+func (p *DeepSeekParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.setInitialState(lastMessage, tools, thinkValue)
+	return tools
+}
+
+type deepseekEvent interface {
+	isDeepSeekEvent()
+}
+
+type deepseekEventThinkingContent struct {
+	content string
+}
+
+type deepseekEventContent struct {
+	content string
+}
+
+type deepseekEventToolCall struct {
+	toolCall api.ToolCall
+}
+
+func (deepseekEventThinkingContent) isDeepSeekEvent() {}
+func (deepseekEventContent) isDeepSeekEvent()         {}
+func (deepseekEventToolCall) isDeepSeekEvent()        {}
+
+func (p *DeepSeekParser) 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 deepseekEventToolCall:
+			toolCalls = append(toolCalls, event.toolCall)
+		case deepseekEventThinkingContent:
+			thinkingSb.WriteString(event.content)
+		case deepseekEventContent:
+			contentSb.WriteString(event.content)
+		}
+	}
+
+	return contentSb.String(), thinkingSb.String(), toolCalls, nil
+}
+
+func (p *DeepSeekParser) parseEvents() []deepseekEvent {
+	var all []deepseekEvent
+
+	keepLooping := true
+	for keepLooping {
+		var events []deepseekEvent
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	return all
+}
+
+func (p *DeepSeekParser) eat() ([]deepseekEvent, bool) {
+	var events []deepseekEvent
+	bufStr := p.buffer.String()
+	if bufStr == "" {
+		return events, false
+	}
+
+	switch p.state {
+	case DeepSeekCollectingThinking:
+		if strings.Contains(bufStr, deepseekThinkingCloseTag) { // thinking[</think>] -> content
+			split := strings.SplitN(bufStr, deepseekThinkingCloseTag, 2)
+			thinking := split[0]
+			thinking = strings.TrimRightFunc(thinking, unicode.IsSpace)
+
+			remaining := split[1]
+			remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = DeepSeekCollectingContent
+
+			if len(thinking) > 0 {
+				events = append(events, deepseekEventThinkingContent{content: thinking})
+			}
+			return events, true
+		} else if overlapLen := overlap(bufStr, deepseekThinkingCloseTag); overlapLen > 0 { // partial </think>
+			beforePartialTag := bufStr[:len(bufStr)-overlapLen]
+			trailingLen := trailingWhitespaceLen(beforePartialTag)
+			ambiguousStart := len(beforePartialTag) - trailingLen
+
+			unambiguous := bufStr[:ambiguousStart]
+			ambiguous := bufStr[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, deepseekEventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		} else { // otherwise its thinking content
+			whitespaceLen := trailingWhitespaceLen(bufStr)
+			ambiguousStart := len(bufStr) - whitespaceLen
+
+			unambiguous := bufStr[:ambiguousStart]
+			ambiguous := bufStr[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, deepseekEventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		}
+
+	case DeepSeekCollectingContent:
+		switch {
+		case strings.Contains(bufStr, deepseekToolCallsBeginTag): // content[<｜tool▁calls▁begin｜>] -> tool calls
+			split := strings.SplitN(bufStr, deepseekToolCallsBeginTag, 2)
+			contentBefore := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			remaining := split[1]
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = DeepSeekCollectingToolCalls
+
+			if len(contentBefore) > 0 {
+				events = append(events, deepseekEventContent{content: contentBefore})
+			}
+			return events, true
+		case strings.Contains(bufStr, deepseekToolOutputBeginTag): // content[<｜tool▁output▁begin｜>] -> tool output
+			split := strings.SplitN(bufStr, deepseekToolOutputBeginTag, 2)
+			contentBefore := split[0] // Don't trim whitespace - preserve spaces
+			remaining := split[1]
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = DeepSeekCollectingToolOutput
+
+			if len(contentBefore) > 0 {
+				events = append(events, deepseekEventContent{content: contentBefore})
+			}
+			return events, true
+		default: // otherwise its content
+			p.buffer.Reset()
+			if len(bufStr) > 0 {
+				events = append(events, deepseekEventContent{content: bufStr})
+			}
+			return events, false
+		}
+
+	case DeepSeekCollectingToolCalls:
+		if idx := strings.Index(bufStr, deepseekToolCallBeginTag); idx != -1 {
+			startIdx := idx + len(deepseekToolCallBeginTag)
+			if endIdx := strings.Index(bufStr[startIdx:], deepseekToolCallEndTag); endIdx != -1 {
+				toolCallContent := bufStr[startIdx : startIdx+endIdx]
+
+				if toolCall, err := p.parseToolCallContent(toolCallContent); err == nil {
+					remaining := bufStr[startIdx+endIdx+len(deepseekToolCallEndTag):]
+					remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+
+					p.buffer.Reset()
+					p.buffer.WriteString(remaining)
+
+					events = append(events, deepseekEventToolCall{toolCall: toolCall})
+					return events, true
+				} else {
+					slog.Warn("deepseek tool call parsing failed", "error", err)
+				}
+			}
+		}
+
+		if idx := strings.Index(bufStr, deepseekToolCallsEndTag); idx != -1 {
+			remaining := bufStr[idx+len(deepseekToolCallsEndTag):]
+			remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = DeepSeekCollectingContent
+
+			return events, true
+		}
+
+		return events, false
+
+	case DeepSeekCollectingToolOutput:
+		if idx := strings.Index(bufStr, deepseekToolOutputEndTag); idx != -1 {
+			toolOutputContent := bufStr[:idx]
+			remaining := bufStr[idx+len(deepseekToolOutputEndTag):]
+			// Don't trim whitespace - preserve spaces after tool output tags
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = DeepSeekCollectingContent
+
+			if len(toolOutputContent) > 0 {
+				events = append(events, deepseekEventContent{content: toolOutputContent})
+			}
+			return events, true
+		}
+
+		return events, false
+	}
+
+	return events, false
+}
+
+func (p *DeepSeekParser) parseToolCallContent(content string) (api.ToolCall, error) {
+	// Expected format: tool_name<｜tool▁sep｜>{args}
+	parts := strings.SplitN(content, deepseekToolSepTag, 2)
+	if len(parts) < 2 {
+		return api.ToolCall{}, errors.New("invalid format")
+	}
+
+	toolName := strings.TrimSpace(parts[0])
+	argsJSON := strings.TrimSpace(parts[1])
+
+	var args api.ToolCallFunctionArguments
+	if err := json.Unmarshal([]byte(argsJSON), &args); err != nil {
+		return api.ToolCall{}, err
+	}
+
+	return api.ToolCall{
+		Function: api.ToolCallFunction{
+			Name:      toolName,
+			Arguments: args,
+		},
+	}, nil
+}

model/parsers/deepseek_test.go

@@ -0,0 +1,721 @@
+package parsers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestDeepSeekParser(t *testing.T) {
+	tests := []struct {
+		name             string
+		input            string
+		expectedContent  string
+		expectedThinking string
+		expectedCalls    []api.ToolCall
+		hasThinking      bool
+	}{
+		{
+			name:            "simple_content",
+			input:           "Hello, how are you?",
+			expectedContent: "Hello, how are you?",
+			hasThinking:     false,
+		},
+		{
+			name:             "thinking_content",
+			input:            "I need to think about this...</think>The answer is 42.",
+			expectedThinking: "I need to think about this...",
+			expectedContent:  "The answer is 42.",
+			hasThinking:      true,
+		},
+		{
+			name:            "no_thinking_simple",
+			input:           "Just a regular response.",
+			expectedContent: "Just a regular response.",
+			hasThinking:     false,
+		},
+		{
+			name:             "thinking_with_newlines",
+			input:            "Let me think:\n- Point 1\n- Point 2</think>\n\nHere's my answer.",
+			expectedThinking: "Let me think:\n- Point 1\n- Point 2",
+			expectedContent:  "Here's my answer.",
+			hasThinking:      true,
+		},
+		{
+			name:            "tool_call_simple",
+			input:           "I'll check the weather.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{\"location\":\"Paris\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedContent: "I'll check the weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "Paris",
+						},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:            "multiple_tool_calls",
+			input:           "Getting weather for both cities.<｜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｜>",
+			expectedContent: "Getting weather for both cities.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "Paris",
+						},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "London",
+						},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:            "tool_output",
+			input:           "Here's the weather: <｜tool▁output▁begin｜>Temperature: 22°C, Sunny<｜tool▁output▁end｜> Hope that helps!",
+			expectedContent: "Here's the weather: Temperature: 22°C, Sunny Hope that helps!",
+			hasThinking:     false,
+		},
+		{
+			name:            "complex_tool_arguments",
+			input:           "Processing data.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>process_data<｜tool▁sep｜>{\"items\":[\"item1\",\"item2\"],\"config\":{\"enabled\":true,\"threshold\":0.95}}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedContent: "Processing data.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "process_data",
+						Arguments: api.ToolCallFunctionArguments{
+							"items":  []interface{}{"item1", "item2"},
+							"config": map[string]interface{}{"enabled": true, "threshold": 0.95},
+						},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:             "thinking_with_tool_call", // technically this can't happen, but the parser can handle it
+			input:            "Let me check the weather...</think>I'll get that for you.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{\"location\":\"Paris\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedThinking: "Let me check the weather...",
+			expectedContent:  "I'll get that for you.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "Paris",
+						},
+					},
+				},
+			},
+			hasThinking: true,
+		},
+		{
+			name:            "empty_content",
+			input:           "",
+			expectedContent: "",
+			hasThinking:     false,
+		},
+		{
+			name:             "only_thinking",
+			input:            "Just thinking content</think>",
+			expectedThinking: "Just thinking content",
+			expectedContent:  "",
+			hasThinking:      true,
+		},
+		{
+			name:            "multiple_tool_outputs",
+			input:           "Results: <｜tool▁output▁begin｜>Paris: 22°C<｜tool▁output▁end｜> and <｜tool▁output▁begin｜>London: 18°C<｜tool▁output▁end｜>",
+			expectedContent: "Results: Paris: 22°C and London: 18°C",
+			hasThinking:     false,
+		},
+		{
+			name:            "unicode_content",
+			input:           "مرحبا بالعالم! 你好世界! 🌍",
+			expectedContent: "مرحبا بالعالم! 你好世界! 🌍",
+			hasThinking:     false,
+		},
+		{
+			name:            "emoji_passthrough",
+			input:           "Task completed ✅ 🎉",
+			expectedContent: "Task completed ✅ 🎉",
+			hasThinking:     false,
+		},
+		{
+			name:            "emoji_after_tool_call",
+			input:           "I'll help you.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{\"location\":\"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>完成 ✅",
+			expectedContent: "I'll help you.完成 ✅",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "Tokyo",
+						},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:            "newlines_and_whitespace",
+			input:           "Line 1\n\nLine 3\t\tTabbed content",
+			expectedContent: "Line 1\n\nLine 3\t\tTabbed content",
+			hasThinking:     false,
+		},
+		{
+			name:             "thinking_with_unicode",
+			input:            "我在思考这个问题...</think>答案是42。",
+			expectedThinking: "我在思考这个问题...",
+			expectedContent:  "答案是42。",
+			hasThinking:      true,
+		},
+		{
+			name:            "tool_call_with_unicode_args",
+			input:           "Searching for information.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>search<｜tool▁sep｜>{\"query\":\"北京天气\",\"language\":\"中文\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedContent: "Searching for information.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "search",
+						Arguments: api.ToolCallFunctionArguments{
+							"query":    "北京天气",
+							"language": "中文",
+						},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:            "tool_output_with_unicode",
+			input:           "天气信息: <｜tool▁output▁begin｜>北京: 25°C, 晴天<｜tool▁output▁end｜> 希望对您有帮助!",
+			expectedContent: "天气信息: 北京: 25°C, 晴天 希望对您有帮助!",
+			hasThinking:     false,
+		},
+		{
+			name:            "mixed_content_with_special_chars",
+			input:           "Price: $100 & tax @ 10% = $110 <｜tool▁output▁begin｜>Total: $110<｜tool▁output▁end｜> (final)",
+			expectedContent: "Price: $100 & tax @ 10% = $110 Total: $110 (final)",
+			hasThinking:     false,
+		},
+		{
+			name:            "tool_call_with_special_chars",
+			input:           "Processing data.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>execute_command<｜tool▁sep｜>{\"command\":\"ls && echo \\\"done\\\"\",\"path\":\"/home/user\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedContent: "Processing data.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "execute_command",
+						Arguments: api.ToolCallFunctionArguments{
+							"command": "ls && echo \"done\"",
+							"path":    "/home/user",
+						},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:             "thinking_with_special_chars",
+			input:            "Let me calculate: 2+2=4 & 3*3=9...</think>The results are correct!",
+			expectedThinking: "Let me calculate: 2+2=4 & 3*3=9...",
+			expectedContent:  "The results are correct!",
+			hasThinking:      true,
+		},
+		{
+			name:            "empty_tool_call_args",
+			input:           "Pinging server.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>ping<｜tool▁sep｜>{}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedContent: "Pinging server.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "ping",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:            "empty_tool_output",
+			input:           "Checking status: <｜tool▁output▁begin｜><｜tool▁output▁end｜> No output received.",
+			expectedContent: "Checking status:  No output received.",
+			hasThinking:     false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			parser := &DeepSeekParser{hasThinkingSupport: tt.hasThinking}
+			parser.Init([]api.Tool{}, nil, &api.ThinkValue{Value: tt.hasThinking})
+
+			content, thinking, calls, err := parser.Add(tt.input, true)
+			if err != nil {
+				t.Fatalf("Add() error = %v", err)
+			}
+
+			if diff := cmp.Diff(tt.expectedContent, content); diff != "" {
+				t.Errorf("Content mismatch (-want +got):\n%s", diff)
+			}
+
+			if diff := cmp.Diff(tt.expectedThinking, thinking); diff != "" {
+				t.Errorf("Thinking mismatch (-want +got):\n%s", diff)
+			}
+
+			if diff := cmp.Diff(tt.expectedCalls, calls); diff != "" {
+				t.Errorf("Tool calls mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestDeepSeekParser_Streaming(t *testing.T) {
+	tests := []struct {
+		name             string
+		chunks           []string
+		expectedContent  string
+		expectedThinking string
+		expectedCalls    []api.ToolCall
+		hasThinking      bool
+	}{
+		{
+			name:            "streaming_simple_content",
+			chunks:          []string{"Hello, ", "how are ", "you?"},
+			expectedContent: "Hello, how are you?",
+			hasThinking:     false,
+		},
+		{
+			name:             "streaming_thinking",
+			chunks:           []string{"I need to ", "think about this", "...</think>", "The answer is 42."},
+			expectedThinking: "I need to think about this...",
+			expectedContent:  "The answer is 42.",
+			hasThinking:      true,
+		},
+		{
+			name:            "streaming_tool_call",
+			chunks:          []string{"I'll check weather.", "<｜tool▁calls▁begin｜>", "<｜tool▁call▁begin｜>get_weather", "<｜tool▁sep｜>{\"location\":\"Paris\"}", "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"},
+			expectedContent: "I'll check weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "Paris",
+						},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:             "streaming_thinking_with_partial_tag",
+			chunks:           []string{"Thinking about this", "...</", "think>", "Done thinking."},
+			expectedThinking: "Thinking about this...",
+			expectedContent:  "Done thinking.",
+			hasThinking:      true,
+		},
+		{
+			name:            "streaming_tool_output",
+			chunks:          []string{"Weather info: ", "<｜tool▁output▁begin｜>", "25°C, Sunny", "<｜tool▁output▁end｜>", " Enjoy!"},
+			expectedContent: "Weather info: 25°C, Sunny Enjoy!",
+			hasThinking:     false,
+		},
+		{
+			name:            "streaming_with_split_tags",
+			chunks:          []string{"Content before ", "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>test", "<｜tool▁sep｜>{}", "<｜tool▁call▁end｜><｜tool▁calls▁end｜>", " after"},
+			expectedContent: "Content before  after",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: api.ToolCallFunctionArguments{},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+		{
+			name:             "streaming_thinking_with_split_end_tag",
+			chunks:           []string{"Thinking content", "</th", "ink>", "Regular content"},
+			expectedThinking: "Thinking content",
+			expectedContent:  "Regular content",
+			hasThinking:      true,
+		},
+		{
+			name:            "streaming_unicode_content",
+			chunks:          []string{"مرحبا ", "بالعالم! ", "你好", "世界!"},
+			expectedContent: "مرحبا بالعالم! 你好世界!",
+			hasThinking:     false,
+		},
+		{
+			name:            "streaming_multiple_tool_outputs",
+			chunks:          []string{"Results: ", "<｜tool▁output▁begin｜>", "Paris: 22°C", "<｜tool▁output▁end｜>", " and ", "<｜tool▁output▁begin｜>", "London: 18°C", "<｜tool▁output▁end｜>"},
+			expectedContent: "Results: Paris: 22°C and London: 18°C",
+			hasThinking:     false,
+		},
+		{
+			name:            "streaming_tool_call_with_split_json",
+			chunks:          []string{"Processing.", "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>calc<｜tool▁sep｜>{\"x\":", "42,\"y\":", "24}<｜tool▁call▁end｜><｜tool▁calls▁end｜>"},
+			expectedContent: "Processing.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "calc",
+						Arguments: api.ToolCallFunctionArguments{
+							"x": float64(42),
+							"y": float64(24),
+						},
+					},
+				},
+			},
+			hasThinking: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			parser := &DeepSeekParser{hasThinkingSupport: tt.hasThinking}
+			parser.Init([]api.Tool{}, nil, &api.ThinkValue{Value: tt.hasThinking})
+
+			var allContent, allThinking string
+			var allCalls []api.ToolCall
+
+			for i, chunk := range tt.chunks {
+				done := i == len(tt.chunks)-1
+				content, thinking, calls, err := parser.Add(chunk, done)
+				if err != nil {
+					t.Fatalf("Add() error = %v", err)
+				}
+
+				allContent += content
+				allThinking += thinking
+				allCalls = append(allCalls, calls...)
+			}
+
+			if diff := cmp.Diff(tt.expectedContent, allContent); diff != "" {
+				t.Errorf("Content mismatch (-want +got):\n%s", diff)
+			}
+
+			if diff := cmp.Diff(tt.expectedThinking, allThinking); diff != "" {
+				t.Errorf("Thinking mismatch (-want +got):\n%s", diff)
+			}
+
+			if diff := cmp.Diff(tt.expectedCalls, allCalls); diff != "" {
+				t.Errorf("Tool calls mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestDeepSeekParser_HasThinkingSupport(t *testing.T) {
+	tests := []struct {
+		name            string
+		hasThinking     bool
+		expectedSupport bool
+	}{
+		{
+			name:            "thinking_enabled",
+			hasThinking:     true,
+			expectedSupport: true,
+		},
+		{
+			name:            "thinking_disabled",
+			hasThinking:     false,
+			expectedSupport: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			parser := &DeepSeekParser{hasThinkingSupport: tt.hasThinking}
+			if got := parser.HasThinkingSupport(); got != tt.expectedSupport {
+				t.Errorf("HasThinkingSupport() = %v, want %v", got, tt.expectedSupport)
+			}
+		})
+	}
+}
+
+func TestDeepSeekParser_HasToolSupport(t *testing.T) {
+	parser := &DeepSeekParser{}
+	if !parser.HasToolSupport() {
+		t.Error("HasToolSupport() should return true")
+	}
+}
+
+func TestDeepSeekParser_Init(t *testing.T) {
+	parser := &DeepSeekParser{hasThinkingSupport: true}
+	tools := []api.Tool{
+		{
+			Type: "function",
+			Function: api.ToolFunction{
+				Name: "test_tool",
+			},
+		},
+	}
+
+	returnedTools := parser.Init(tools, nil, &api.ThinkValue{Value: true})
+
+	if diff := cmp.Diff(tools, returnedTools); diff != "" {
+		t.Errorf("Init() returned tools mismatch (-want +got):\n%s", diff)
+	}
+
+	// Test initial state is set to thinking when enabled
+	if parser.state != DeepSeekCollectingThinking {
+		t.Errorf("Expected initial state to be DeepSeekCollectingThinking, got %v", parser.state)
+	}
+}
+
+func TestDeepSeekParser_parseToolCallContent(t *testing.T) {
+	tests := []struct {
+		name        string
+		content     string
+		expected    api.ToolCall
+		expectError bool
+	}{
+		{
+			name:    "valid_tool_call",
+			content: "get_weather<｜tool▁sep｜>{\"location\":\"Paris\"}",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "get_weather",
+					Arguments: api.ToolCallFunctionArguments{
+						"location": "Paris",
+					},
+				},
+			},
+		},
+		{
+			name:    "complex_arguments",
+			content: "process_data<｜tool▁sep｜>{\"items\":[\"a\",\"b\"],\"config\":{\"enabled\":true}}",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "process_data",
+					Arguments: api.ToolCallFunctionArguments{
+						"items":  []interface{}{"a", "b"},
+						"config": map[string]interface{}{"enabled": true},
+					},
+				},
+			},
+		},
+		{
+			name:    "empty_arguments",
+			content: "ping<｜tool▁sep｜>{}",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "ping",
+					Arguments: api.ToolCallFunctionArguments{},
+				},
+			},
+		},
+		{
+			name:    "unicode_in_tool_name",
+			content: "获取天气<｜tool▁sep｜>{\"城市\":\"北京\"}",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "获取天气",
+					Arguments: api.ToolCallFunctionArguments{
+						"城市": "北京",
+					},
+				},
+			},
+		},
+		{
+			name:    "special_chars_in_arguments",
+			content: "execute<｜tool▁sep｜>{\"command\":\"ls && echo \\\"done\\\"\",\"path\":\"/home/user\"}",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "execute",
+					Arguments: api.ToolCallFunctionArguments{
+						"command": "ls && echo \"done\"",
+						"path":    "/home/user",
+					},
+				},
+			},
+		},
+		{
+			name:    "numeric_arguments",
+			content: "calculate<｜tool▁sep｜>{\"x\":3.14,\"y\":42,\"enabled\":true}",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "calculate",
+					Arguments: api.ToolCallFunctionArguments{
+						"x":       3.14,
+						"y":       float64(42),
+						"enabled": true,
+					},
+				},
+			},
+		},
+		{
+			name:        "invalid_format_no_separator",
+			content:     "get_weather{\"location\":\"Paris\"}",
+			expectError: true,
+		},
+		{
+			name:        "invalid_json",
+			content:     "get_weather<｜tool▁sep｜>{invalid json}",
+			expectError: true,
+		},
+		{
+			name:        "empty_tool_name",
+			content:     "<｜tool▁sep｜>{\"arg\":\"value\"}",
+			expectError: false, // This should work, just empty name
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "",
+					Arguments: api.ToolCallFunctionArguments{
+						"arg": "value",
+					},
+				},
+			},
+		},
+		{
+			name:        "missing_json_part",
+			content:     "tool_name<｜tool▁sep｜>",
+			expectError: true,
+		},
+	}
+
+	parser := &DeepSeekParser{}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result, err := parser.parseToolCallContent(tt.content)
+
+			if tt.expectError {
+				if err == nil {
+					t.Error("Expected error but got none")
+				}
+				return
+			}
+
+			if err != nil {
+				t.Fatalf("Unexpected error: %v", err)
+			}
+
+			if diff := cmp.Diff(tt.expected, result); diff != "" {
+				t.Errorf("parseToolCallContent() mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestDeepSeekParser_EdgeCases(t *testing.T) {
+	tests := []struct {
+		name             string
+		input            string
+		expectedContent  string
+		expectedThinking string
+		hasThinking      bool
+	}{
+		{
+			name:             "nested_think_tags_in_thinking",
+			input:            "Outer thinking <think>inner</think> content</think>Final content",
+			expectedThinking: "Outer thinking <think>inner",
+			expectedContent:  "content</think>Final content",
+			hasThinking:      true,
+		},
+		{
+			name:             "multiple_think_close_tags",
+			input:            "First thought</think>Second thought</think>Final content",
+			expectedThinking: "First thought",
+			expectedContent:  "Second thought</think>Final content",
+			hasThinking:      true,
+		},
+		{
+			name:             "empty_thinking_content",
+			input:            "</think>Just content",
+			expectedThinking: "",
+			expectedContent:  "Just content",
+			hasThinking:      true,
+		},
+		{
+			name:            "thinking_disabled_with_think_tags",
+			input:           "Some content</think>More content",
+			expectedContent: "Some content</think>More content",
+			hasThinking:     false,
+		},
+		{
+			name:            "malformed_tool_call_missing_sep",
+			input:           "Testing.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>bad_tool{\"arg\":\"value\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedContent: "Testing.",
+			hasThinking:     false,
+		},
+		{
+			name:            "malformed_tool_call_invalid_json",
+			input:           "Testing.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>bad_tool<｜tool▁sep｜>{invalid json}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedContent: "Testing.",
+			hasThinking:     false,
+		},
+		{
+			name:            "partial_tool_tag_at_end",
+			input:           "Content with partial <｜tool▁calls▁",
+			expectedContent: "Content with partial <｜tool▁calls▁",
+			hasThinking:     false,
+		},
+		{
+			name:            "partial_think_tag_at_end",
+			input:           "Thinking content</th",
+			expectedContent: "Thinking content</th",
+			hasThinking:     false,
+		},
+		{
+			name:             "partial_think_tag_at_end_with_thinking",
+			input:            "Thinking content</th",
+			expectedThinking: "Thinking content",
+			expectedContent:  "",
+			hasThinking:      true,
+		},
+		{
+			name:            "whitespace_only_content",
+			input:           "   \n\t   ",
+			expectedContent: "   \n\t   ",
+			hasThinking:     false,
+		},
+		{
+			name:            "tool_output_with_newlines",
+			input:           "Output:\n<｜tool▁output▁begin｜>Line 1\nLine 2\nLine 3<｜tool▁output▁end｜>\nDone.",
+			expectedContent: "Output:\nLine 1\nLine 2\nLine 3\nDone.",
+			hasThinking:     false,
+		},
+		{
+			name:            "consecutive_tool_calls",
+			input:           "First.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool1<｜tool▁sep｜>{}<｜tool▁call▁end｜><｜tool▁calls▁end｜>Second.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>tool2<｜tool▁sep｜>{}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedContent: "First.",
+			hasThinking:     false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			parser := &DeepSeekParser{hasThinkingSupport: tt.hasThinking}
+			parser.Init([]api.Tool{}, nil, &api.ThinkValue{Value: tt.hasThinking})
+
+			content, thinking, _, err := parser.Add(tt.input, true)
+			if err != nil {
+				t.Fatalf("Add() error = %v", err)
+			}
+
+			if diff := cmp.Diff(tt.expectedContent, content); diff != "" {
+				t.Errorf("Content mismatch (-want +got):\n%s", diff)
+			}
+
+			if diff := cmp.Diff(tt.expectedThinking, thinking); diff != "" {
+				t.Errorf("Thinking mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}

model/parsers/parsers.go

@@ -58,6 +58,8 @@ func ParserForName(name string) Parser {
 		return harmony.NewHarmonyMessageHandler()
 	case "cogito":
 		return &CogitoParser{}
+	case "deepseek":
+		return &DeepSeekParser{hasThinkingSupport: true}
 	case "olmo3":
 		return &Olmo3Parser{}
 	case "olmo3-think":

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

@@ -60,10 +60,18 @@ func rendererForName(name string) Renderer {
 		renderer := &CogitoRenderer{isThinking: true}
 		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
 	default:
 		return nil