linter: remove prealloc

The cuda_jetpack libs will enumerate discrete GPUs on SBSA systems
which leads to runtime failures of missing kernels.  This fix
requires an exact match to enable jetpacks instead of relying on
enumeration to filter out supported libraries.

.gitattributes

@@ -15,8 +15,12 @@ ml/backend/**/*.cu linguist-vendored
 ml/backend/**/*.cuh linguist-vendored
 ml/backend/**/*.m linguist-vendored
 ml/backend/**/*.metal linguist-vendored
+ml/backend/**/*.comp linguist-vendored
+ml/backend/**/*.glsl linguist-vendored
 ml/backend/**/CMakeLists.txt linguist-vendored
 
+app/webview linguist-vendored
+
 llama/build-info.cpp linguist-generated
 ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.s linguist-generated
 

.golangci.yaml

@@ -23,7 +23,6 @@ linters:
     - nolintlint
     - nosprintfhostport
     - perfsprint
-    - prealloc
     - sloglint
     - staticcheck
     - unconvert

app/cmd/app/app.go

@@ -397,8 +397,8 @@ func checkUserLoggedIn(uiServerPort int) bool {
 // handleConnectURLScheme fetches the connect URL and opens it in the browser
 func handleConnectURLScheme() {
 	if checkUserLoggedIn(uiServerPort) {
-		slog.Info("user is already logged in, opening settings instead")
-		sendUIRequestMessage("/")
+		slog.Info("user is already logged in, opening app instead")
+		showWindow(wv.webview.Window())
 		return
 	}
 
@@ -466,6 +466,8 @@ func handleURLSchemeInCurrentInstance(urlSchemeRequest string) {
 	if isConnect {
 		handleConnectURLScheme()
 	} else {
-		sendUIRequestMessage("/")
+		if wv.webview != nil {
+			showWindow(wv.webview.Window())
+		}
 	}
 }

app/cmd/app/app_darwin.m

@@ -24,27 +24,14 @@ bool firstTimeRun,startHidden; // Set in run before initialization
     for (NSURL *url in urls) {
         if ([url.scheme isEqualToString:@"ollama"]) {
             NSString *path = url.path;
-            if (!path || [path isEqualToString:@""]) {
-                // For URLs like ollama://settings (without triple slash),
-                // the "settings" part is parsed as the host, not the path.
-                // We need to convert it to a path by prepending "/"
-                if (url.host && ![url.host isEqualToString:@""]) {
-                    path = [@"/" stringByAppendingString:url.host];
-                } else {
-                    path = @"/";
-                }
-            }
-            
-            if ([path isEqualToString:@"/connect"] || [url.host isEqualToString:@"connect"]) {
+
+            if (path && ([path isEqualToString:@"/connect"] || [url.host isEqualToString:@"connect"])) {
                 // Special case: handle connect by opening browser instead of app
                 handleConnectURL();
             } else {
                 // Set app to be active and visible
                 [NSApp setActivationPolicy:NSApplicationActivationPolicyRegular];
                 [NSApp activateIgnoringOtherApps:YES];
-                
-                // Open the path with the UI
-                [self uiRequest:path];
             }
             
             break;
@@ -260,7 +247,7 @@ bool firstTimeRun,startHidden; // Set in run before initialization
 }
 
 - (void)openHelp:(id)sender {
-    NSURL *url = [NSURL URLWithString:@"https://github.com/ollama/ollama/tree/main/docs"];
+    NSURL *url = [NSURL URLWithString:@"https://docs.ollama.com/"];
     [[NSWorkspace sharedWorkspace] openURL:url];
 }
 

app/cmd/app/app_windows.go

@@ -147,7 +147,9 @@ func handleURLSchemeRequest(urlScheme string) {
 	if isConnect {
 		handleConnectURLScheme()
 	} else {
-		sendUIRequestMessage("/")
+		if wv.webview != nil {
+			showWindow(wv.webview.Window())
+		}
 	}
 }
 

discover/runner.go

@@ -98,6 +98,9 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 					continue
 				} else if jetpack != "" && filepath.Base(dir) != "cuda_"+jetpack {
 					continue
+				} else if jetpack == "" && strings.Contains(filepath.Base(dir), "cuda_jetpack") {
+					slog.Debug("jetpack not detected (set JETSON_JETPACK or OLLAMA_LLM_LIBRARY to override), skipping", "libDir", dir)
+					continue
 				} else if !envconfig.EnableVulkan() && strings.Contains(filepath.Base(dir), "vulkan") {
 					slog.Info("experimental Vulkan support disabled.  To enable, set OLLAMA_VULKAN=1")
 					continue
@@ -125,10 +128,20 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 		supportedMu := sync.Mutex{}
 		supported := make(map[string]map[string]map[string]int) // [Library][libDir][ID] = pre-deletion devices index
 		for i := range devices {
+			libDir := devices[i].LibraryPath[len(devices[i].LibraryPath)-1]
 			if !devices[i].NeedsInitValidation() {
+				// No need to validate, add to the supported map
+				supportedMu.Lock()
+				if _, ok := supported[devices[i].Library]; !ok {
+					supported[devices[i].Library] = make(map[string]map[string]int)
+				}
+				if _, ok := supported[devices[i].Library][libDir]; !ok {
+					supported[devices[i].Library][libDir] = make(map[string]int)
+				}
+				supported[devices[i].Library][libDir][devices[i].ID] = i
+				supportedMu.Unlock()
 				continue
 			}
-			libDir := devices[i].LibraryPath[len(devices[i].LibraryPath)-1]
 			slog.Debug("verifying if device is supported", "library", libDir, "description", devices[i].Description, "compute", devices[i].Compute(), "id", devices[i].ID, "pci_id", devices[i].PCIID)
 			wg.Add(1)
 			go func(i int) {

docs/faq.mdx

@@ -57,8 +57,13 @@ ollama ps
 ```
 
 <Info>
-  **Output**: ``` NAME ID SIZE PROCESSOR UNTIL llama3:70b bcfb190ca3a7 42 GB
-  100% GPU 4 minutes from now ```
+
+**Output**:
+
+```
+NAME        ID            SIZE    PROCESSOR   UNTIL
+llama3:70b  bcfb190ca3a7  42 GB   100% GPU    4 minutes from now
+```
 </Info>
 
 The `Processor` column will show which memory the model was loaded in to:
@@ -385,4 +390,4 @@ Ollama for Windows and macOS register as a login item during installation.  You
 - In `Task Manager` go to the `Startup apps` tab, search for `ollama` then click `Disable`
 
 **MacOS**
-- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.
+- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.

docs/modelfile.mdx

@@ -149,9 +149,6 @@ PARAMETER <parameter> <parametervalue>
 
 | Parameter      | Description                                                                                                                                                                                                                                                                                                                                                                     | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
-| mirostat       | Enable Mirostat sampling for controlling perplexity. (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)                                                                                                                                                                                                                                                                 | int        | mirostat 0           |
-| mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                                                                                                                                                | float      | mirostat_eta 0.1     |
-| mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                                                                                                                                                 | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                                                                                                                                            | int        | num_ctx 4096         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                                                                                                                                                   | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                                                                                                                                             | float      | repeat_penalty 1.1   |

fs/ggml/ggml.go

@@ -251,6 +251,7 @@ func (kv KV) OllamaEngineRequired() bool {
 		"qwen3vl", "qwen3vlmoe",
 		"deepseekocr",
 		"deepseek2",
+		"nomic-bert",
 	}, kv.Architecture())
 }
 

harmony/harmonyparser.go

@@ -388,9 +388,9 @@ func NewFunctionNameMap() *FunctionNameMap {
 	}
 }
 
-// Init initializes the handler with tools and optional last message
+// Init initializes the handler with tools, optional last message, and think value
 // Implements the Parser interface
-func (h *HarmonyMessageHandler) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
+func (h *HarmonyMessageHandler) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
 	// Initialize the harmony parser
 	if h.HarmonyParser == nil {
 		h.HarmonyParser = &HarmonyParser{

model/models/deepseek2/model.go

@@ -236,11 +236,6 @@ type Model struct {
 }
 
 func New(c fs.Config) (model.Model, error) {
-	if c.Uint("attention.key_length_mla") == 0 {
-		// non-MLA models aren't yet supported
-		return nil, model.ErrUnsupportedModel
-	}
-
 	layers := make([]Layer, c.Uint("block_count"))
 
 	firstDenseLayerIndex := int(c.Uint("leading_dense_block_count"))
@@ -259,6 +254,30 @@ func New(c fs.Config) (model.Model, error) {
 	keyLength := int(cmp.Or(c.Uint("attention.key_length_mla"), c.Uint("attention.key_length")))
 	valueLength := int(cmp.Or(c.Uint("attention.value_length_mla"), c.Uint("attention.value_length")))
 
+	var pre []string
+	switch c.String("tokenizer.ggml.pre") {
+	case "deepseek-v3":
+		pre = []string{
+			// Split regex into multiple parts (according to DeepSeek3's regex)
+			"\\p{N}{1,3}",
+			`[一-龥぀-ゟ゠-ヿ]+`,
+			"[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\r\n]*|\\s*[\r\n]+|\\s+(?!\\S)|\\s+",
+		}
+	case "deepseek-llm":
+		// TODO: these models haven't been vetted so skip for now
+		// pre = []string{
+		// 	"[\r\n]",
+		// 	"\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
+		// 	"\\s?[!-/:-~！-／：-～‘-‟　-。]+",
+		// 	"\\s+$",
+		// 	"[一-龥ࠀ-一가-퟿]+",
+		// 	"[0-9]",
+		// }
+		fallthrough
+	default:
+		return nil, model.ErrUnsupportedTokenizer
+	}
+
 	m := Model{
 		BytePairEncoding: model.NewBytePairEncoding(
 			&model.Vocabulary{
@@ -273,10 +292,7 @@ func New(c fs.Config) (model.Model, error) {
 					c.Ints("tokenizer.ggml.eos_token_ids")...,
 				),
 			},
-			// Split regex into multiple parts (according to DeepSeek3's regex)
-			"\\p{N}{1,3}",
-			`[一-龥぀-ゟ゠-ヿ]+`,
-			"[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\r\n]*|\\s*[\r\n]+|\\s+(?!\\S)|\\s+",
+			pre...,
 		),
 		Layers: layers,
 		Options: &Options{

model/parsers/cogito.go

@@ -0,0 +1,319 @@
+package parsers
+
+import (
+	"encoding/json"
+	"errors"
+	"log/slog"
+	"strings"
+	"unicode"
+
+	"github.com/ollama/ollama/api"
+)
+
+type CogitoParserState int
+
+const (
+	CogitoCollectingThinking CogitoParserState = iota
+	CogitoCollectingContent
+	CogitoCollectingToolCalls
+	CogitoCollectingToolOutput
+)
+
+const (
+	cogitoThinkingCloseTag    = "</think>"
+	cogitoToolCallsBeginTag   = "<｜tool▁calls▁begin｜>"
+	cogitoToolCallsEndTag     = "<｜tool▁calls▁end｜>"
+	cogitoToolCallBeginTag    = "<｜tool▁call▁begin｜>"
+	cogitoToolCallEndTag      = "<｜tool▁call▁end｜>"
+	cogitoToolSepTag          = "<｜tool▁sep｜>"
+	cogitoToolOutputBeginTag  = "<｜tool▁output▁begin｜>"
+	cogitoToolOutputEndTag    = "<｜tool▁output▁end｜>"
+	cogitoToolOutputsBeginTag = "<｜tool▁outputs▁begin｜>"
+	cogitoToolOutputsEndTag   = "<｜tool▁outputs▁end｜>"
+)
+
+type CogitoParser struct {
+	state  CogitoParserState
+	buffer strings.Builder
+}
+
+func (p *CogitoParser) HasToolSupport() bool {
+	return true
+}
+
+func (p *CogitoParser) HasThinkingSupport() bool {
+	return true
+}
+
+func (p *CogitoParser) setInitialState(lastMessage *api.Message, tools []api.Tool, thinkValue *api.ThinkValue) {
+	prefill := lastMessage != nil && lastMessage.Role == "assistant"
+
+	// Check both model capability AND request preference
+	thinkingEnabled := thinkValue != nil && thinkValue.Bool()
+	// thinkingEnabled should be set to false for tools
+
+	if !thinkingEnabled {
+		p.state = CogitoCollectingContent
+		return
+	}
+
+	if prefill && lastMessage.Content != "" {
+		p.state = CogitoCollectingContent
+		return
+	}
+
+	// Note: for cogito, if there are tools, then we don't want to be thinking
+	if len(tools) > 0 {
+		p.state = CogitoCollectingContent
+		return
+	}
+
+	p.state = CogitoCollectingThinking
+}
+
+func (p *CogitoParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.setInitialState(lastMessage, tools, thinkValue)
+	return tools
+}
+
+type cogitoEvent interface {
+	isCogitoEvent()
+}
+
+type cogitoEventThinkingContent struct {
+	content string
+}
+
+type cogitoEventContent struct {
+	content string
+}
+
+type cogitoEventToolCall struct {
+	toolCall api.ToolCall
+}
+
+func (cogitoEventThinkingContent) isCogitoEvent() {}
+func (cogitoEventContent) isCogitoEvent()         {}
+func (cogitoEventToolCall) isCogitoEvent()        {}
+
+func (p *CogitoParser) 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 cogitoEventToolCall:
+			toolCalls = append(toolCalls, event.toolCall)
+		case cogitoEventThinkingContent:
+			thinkingSb.WriteString(event.content)
+		case cogitoEventContent:
+			contentSb.WriteString(event.content)
+		}
+	}
+
+	return contentSb.String(), thinkingSb.String(), toolCalls, nil
+}
+
+func (p *CogitoParser) parseEvents() []cogitoEvent {
+	var all []cogitoEvent
+
+	keepLooping := true
+	for keepLooping {
+		var events []cogitoEvent
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	return all
+}
+
+func (p *CogitoParser) eat() ([]cogitoEvent, bool) {
+	var events []cogitoEvent
+	bufStr := p.buffer.String()
+	if bufStr == "" {
+		return events, false
+	}
+
+	switch p.state {
+	case CogitoCollectingThinking:
+		if strings.Contains(bufStr, cogitoThinkingCloseTag) { // thinking[</think>] -> content
+			split := strings.SplitN(bufStr, cogitoThinkingCloseTag, 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 = CogitoCollectingContent
+
+			if len(thinking) > 0 {
+				events = append(events, cogitoEventThinkingContent{content: thinking})
+			}
+			return events, true
+		} else if overlapLen := overlap(bufStr, cogitoThinkingCloseTag); 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, cogitoEventThinkingContent{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, cogitoEventThinkingContent{content: unambiguous})
+			}
+			return events, false
+		}
+
+	case CogitoCollectingContent:
+		switch {
+		case strings.Contains(bufStr, cogitoToolCallsBeginTag): // content[<｜tool▁calls▁begin｜>] -> tool calls
+			split := strings.SplitN(bufStr, cogitoToolCallsBeginTag, 2)
+			contentBefore := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			remaining := split[1]
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = CogitoCollectingToolCalls
+
+			if len(contentBefore) > 0 {
+				events = append(events, cogitoEventContent{content: contentBefore})
+			}
+			return events, true
+		case strings.Contains(bufStr, cogitoToolOutputsBeginTag): // content[<｜tool▁outputs▁begin｜>] -> tool outputs
+			split := strings.SplitN(bufStr, cogitoToolOutputsBeginTag, 2)
+			contentBefore := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			remaining := split[1]
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = CogitoCollectingToolOutput
+
+			if len(contentBefore) > 0 {
+				events = append(events, cogitoEventContent{content: contentBefore})
+			}
+			return events, true
+		default: // otherwise its content
+			p.buffer.Reset()
+			if len(bufStr) > 0 {
+				events = append(events, cogitoEventContent{content: bufStr})
+			}
+			return events, false
+		}
+	case CogitoCollectingToolCalls:
+		if idx := strings.Index(bufStr, cogitoToolCallBeginTag); idx != -1 {
+			startIdx := idx + len(cogitoToolCallBeginTag)
+			if endIdx := strings.Index(bufStr[startIdx:], cogitoToolCallEndTag); endIdx != -1 {
+				toolCallContent := bufStr[startIdx : startIdx+endIdx]
+
+				if toolCall, err := p.parseToolCallContent(toolCallContent); err == nil {
+					remaining := bufStr[startIdx+endIdx+len(cogitoToolCallEndTag):]
+					remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+
+					p.buffer.Reset()
+					p.buffer.WriteString(remaining)
+
+					events = append(events, cogitoEventToolCall{toolCall: toolCall})
+					return events, true
+				} else {
+					slog.Warn("cogito tool call parsing failed", "error", err)
+				}
+			}
+		}
+
+		if idx := strings.Index(bufStr, cogitoToolCallsEndTag); idx != -1 {
+			remaining := bufStr[idx+len(cogitoToolCallsEndTag):]
+			remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = CogitoCollectingContent
+
+			return events, true
+		}
+
+		return events, false
+
+	case CogitoCollectingToolOutput:
+		if idx := strings.Index(bufStr, cogitoToolOutputBeginTag); idx != -1 {
+			startIdx := idx + len(cogitoToolOutputBeginTag)
+			if endIdx := strings.Index(bufStr[startIdx:], cogitoToolOutputEndTag); endIdx != -1 {
+				remaining := bufStr[startIdx+endIdx+len(cogitoToolOutputEndTag):]
+				remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+
+				p.buffer.Reset()
+				p.buffer.WriteString(remaining)
+
+				return events, true
+			}
+		}
+
+		if idx := strings.Index(bufStr, cogitoToolOutputsEndTag); idx != -1 {
+			remaining := bufStr[idx+len(cogitoToolOutputsEndTag):]
+			remaining = strings.TrimLeftFunc(remaining, unicode.IsSpace)
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = CogitoCollectingContent
+
+			return events, true
+		}
+
+		return events, false
+	}
+
+	return events, false
+}
+
+func (p *CogitoParser) parseToolCallContent(content string) (api.ToolCall, error) {
+	// Expected format: function<｜tool▁sep｜>tool_name\n```json\n{args}\n```
+	parts := strings.SplitN(content, cogitoToolSepTag, 2)
+	if len(parts) < 2 {
+		return api.ToolCall{}, errors.New("invalid format")
+	}
+	nameAndArgs := parts[1]
+
+	jsonStart := strings.Index(nameAndArgs, "\n```json\n")
+	if jsonStart == -1 {
+		return api.ToolCall{}, errors.New("invalid format")
+	}
+	toolName := strings.TrimSpace(nameAndArgs[:jsonStart])
+	jsonContent := nameAndArgs[jsonStart+len("\n```json\n"):]
+
+	jsonEnd := strings.Index(jsonContent, "\n```")
+	if jsonEnd == -1 {
+		return api.ToolCall{}, errors.New("invalid format")
+	}
+	argsJSON := jsonContent[:jsonEnd]
+
+	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/cogito_test.go

@@ -0,0 +1,565 @@
+package parsers
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestCogitoParser(t *testing.T) {
+	tests := []struct {
+		name              string
+		input             string
+		expectedContent   string
+		expectedThinking  string
+		expectedToolCalls []api.ToolCall
+		tools             []api.Tool
+		lastMessage       *api.Message
+	}{
+		{
+			name:             "simple_content",
+			input:            "This is a simple response.",
+			expectedContent:  "This is a simple response.",
+			expectedThinking: "",
+		},
+		{
+			name:             "thinking_only",
+			input:            "This is thinking content.</think>This is response content.",
+			expectedContent:  "This is response content.",
+			expectedThinking: "This is thinking content.",
+		},
+		{
+			name: "tool_call_simple",
+			input: `<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather
+` + "```json\n" + `{"location":"Paris"}
+` + "```" + `<｜tool▁call▁end｜><｜tool▁calls▁end｜>`,
+			expectedToolCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "Paris",
+						},
+					},
+				},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "get_weather",
+						Parameters: api.ToolFunctionParameters{
+							Properties: map[string]api.ToolProperty{
+								"location": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "thinking_with_tool_call",
+			input: `I need to check the weather.</think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather
+` + "```json\n" + `{"location":"Paris"}
+` + "```" + `<｜tool▁call▁end｜><｜tool▁calls▁end｜>`,
+			expectedContent:  "I need to check the weather.</think>",
+			expectedThinking: "", // No thinking when tools are present (Cogito-specific behavior)
+			expectedToolCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "Paris",
+						},
+					},
+				},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "get_weather",
+						Parameters: api.ToolFunctionParameters{
+							Properties: map[string]api.ToolProperty{
+								"location": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "multiple_tool_calls",
+			input: `<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather
+` + "```json\n" + `{"location":"Paris"}
+` + "```" + `<｜tool▁call▁end｜>
+<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather
+` + "```json\n" + `{"location":"London"}
+` + "```" + `<｜tool▁call▁end｜><｜tool▁calls▁end｜>`,
+			expectedToolCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "Paris",
+						},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name: "get_weather",
+						Arguments: api.ToolCallFunctionArguments{
+							"location": "London",
+						},
+					},
+				},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "get_weather",
+						Parameters: api.ToolFunctionParameters{
+							Properties: map[string]api.ToolProperty{
+								"location": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "complex_tool_arguments",
+			input: `<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>process_data
+` + "```json\n" + `{"items":["item1","item2"],"config":{"enabled":true,"threshold":0.95},"count":42}
+` + "```" + `<｜tool▁call▁end｜><｜tool▁calls▁end｜>`,
+			expectedToolCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "process_data",
+						Arguments: api.ToolCallFunctionArguments{
+							"items":  []any{"item1", "item2"},
+							"config": map[string]any{"enabled": true, "threshold": 0.95},
+							"count":  42.0,
+						},
+					},
+				},
+			},
+		},
+		{
+			name:             "tool_output_parsing",
+			input:            `<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>{"temperature": 22, "condition": "sunny"}<｜tool▁output▁end｜><｜tool▁outputs▁end｜>`,
+			expectedContent:  "",
+			expectedThinking: "",
+		},
+		{
+			name: "thinking_with_multiline_content",
+			input: `This is line 1
+This is line 2
+This is line 3</think>Final response here.`,
+			expectedContent:  "Final response here.",
+			expectedThinking: "This is line 1\nThis is line 2\nThis is line 3",
+		},
+		{
+			name:             "no_thinking_simple",
+			input:            "This is content.",
+			expectedContent:  "This is content.",
+			expectedThinking: "",
+		},
+		{
+			name:            "prefill_content_only",
+			input:           "Continuing from previous content.",
+			expectedContent: "Continuing from previous content.",
+			lastMessage: &api.Message{
+				Role:    "assistant",
+				Content: "Previous content",
+			},
+		},
+		{
+			name:             "prefill_with_thinking",
+			input:            "Continuing thinking</think>Continuing content.",
+			expectedContent:  "Continuing content.",
+			expectedThinking: "Continuing thinking",
+			lastMessage: &api.Message{
+				Role: "assistant",
+			},
+		},
+		// Edge cases
+		{
+			name:             "nested_think_tags_in_thinking",
+			input:            "I'm thinking <think>nested</think> more thinking</think>Final content.",
+			expectedContent:  "more thinking</think>Final content.",
+			expectedThinking: "I'm thinking <think>nested",
+		},
+		{
+			name:             "multiple_think_close_tags",
+			input:            "First thinking</think>Content</think>More content.",
+			expectedContent:  "Content</think>More content.",
+			expectedThinking: "First thinking",
+		},
+		{
+			name:             "empty_thinking_content",
+			input:            "</think>Just content here.",
+			expectedContent:  "</think>Just content here.",
+			expectedThinking: "",
+		},
+		{
+			name:             "thinking_disabled_with_think_tags",
+			input:            "Content with </think> tags should be treated as content.",
+			expectedContent:  "Content with </think> tags should be treated as content.",
+			expectedThinking: "",
+			lastMessage: &api.Message{
+				Role:    "assistant",
+				Content: "existing", // Forces non-thinking mode
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			// Use thinking-enabled parser for tests that expect thinking
+			hasThinking := tt.expectedThinking != ""
+			parser := &CogitoParser{}                                                  // it has thinking support
+			parser.Init(tt.tools, tt.lastMessage, &api.ThinkValue{Value: hasThinking}) // but we should set it with the request that the user wants
+
+			content, thinking, toolCalls, 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.expectedToolCalls, toolCalls); diff != "" {
+				t.Errorf("tool calls mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestCogitoParser_Streaming(t *testing.T) {
+	parser := &CogitoParser{}
+	parser.Init(nil, nil, &api.ThinkValue{Value: true})
+
+	chunks := []string{
+		"This is ",
+		"thinking content",
+		".</think>This is ",
+		"content.<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>test_tool\n```json\n{\"arg\":\"value\"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+	}
+
+	var finalContent, finalThinking strings.Builder
+	var finalToolCalls []api.ToolCall
+
+	for i, chunk := range chunks {
+		done := i == len(chunks)-1
+		content, thinking, toolCalls, err := parser.Add(chunk, done)
+		if err != nil {
+			t.Fatalf("Add() error on chunk %d: %v", i, err)
+		}
+
+		finalContent.WriteString(content)
+		finalThinking.WriteString(thinking)
+		finalToolCalls = append(finalToolCalls, toolCalls...)
+	}
+
+	expectedContent := "This is content."
+	expectedThinking := "This is thinking content."
+	expectedToolCalls := []api.ToolCall{
+		{
+			Function: api.ToolCallFunction{
+				Name: "test_tool",
+				Arguments: api.ToolCallFunctionArguments{
+					"arg": "value",
+				},
+			},
+		},
+	}
+
+	if finalContent.String() != expectedContent {
+		t.Errorf("expected content %q, got %q", expectedContent, finalContent.String())
+	}
+
+	if finalThinking.String() != expectedThinking {
+		t.Errorf("expected thinking %q, got %q", expectedThinking, finalThinking.String())
+	}
+
+	if diff := cmp.Diff(expectedToolCalls, finalToolCalls); diff != "" {
+		t.Errorf("tool calls mismatch (-want +got):\n%s", diff)
+	}
+}
+
+func TestCogitoParser_StreamingEdgeCases(t *testing.T) {
+	tests := []struct {
+		name               string
+		chunks             []string
+		expectedContent    string
+		expectedThinking   string
+		expectedToolCalls  []api.ToolCall
+		hasThinkingSupport bool
+	}{
+		{
+			name: "split_thinking_tag",
+			chunks: []string{
+				"This is thinking content</thi",
+				"nk>This is content.",
+			},
+			expectedContent:    "This is content.",
+			expectedThinking:   "This is thinking content",
+			hasThinkingSupport: true,
+		},
+		{
+			name: "split_tool_calls_begin_tag_conservative_parsing",
+			chunks: []string{
+				"Content before<｜tool▁calls▁beg",
+				"in｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>test\n```json\n{}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			},
+			// Parser is conservative - treats incomplete tags as content
+			expectedContent:    "Content before<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>test\n```json\n{}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+			expectedToolCalls:  nil,
+			hasThinkingSupport: false,
+		},
+		{
+			name: "thinking_disabled_with_split_tags",
+			chunks: []string{
+				"Content with </thi",
+				"nk> should be treated as content.",
+			},
+			expectedContent:    "Content with </think> should be treated as content.",
+			expectedThinking:   "",
+			hasThinkingSupport: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			parser := &CogitoParser{}
+			parser.Init(nil, nil, &api.ThinkValue{Value: tt.hasThinkingSupport})
+
+			var finalContent, finalThinking strings.Builder
+			var finalToolCalls []api.ToolCall
+
+			for i, chunk := range tt.chunks {
+				done := i == len(tt.chunks)-1
+				content, thinking, toolCalls, err := parser.Add(chunk, done)
+				if err != nil {
+					t.Fatalf("Add() error on chunk %d: %v", i, err)
+				}
+
+				finalContent.WriteString(content)
+				finalThinking.WriteString(thinking)
+				finalToolCalls = append(finalToolCalls, toolCalls...)
+			}
+
+			if finalContent.String() != tt.expectedContent {
+				t.Errorf("expected content %q, got %q", tt.expectedContent, finalContent.String())
+			}
+
+			if finalThinking.String() != tt.expectedThinking {
+				t.Errorf("expected thinking %q, got %q", tt.expectedThinking, finalThinking.String())
+			}
+
+			if diff := cmp.Diff(tt.expectedToolCalls, finalToolCalls); diff != "" {
+				t.Errorf("tool calls mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestCogitoParser_HasToolSupport(t *testing.T) {
+	parser := &CogitoParser{}
+	if !parser.HasToolSupport() {
+		t.Error("CogitoParser should support tools")
+	}
+}
+
+func TestCogitoParser_Init(t *testing.T) {
+	parser := &CogitoParser{}
+
+	tools := []api.Tool{
+		{Function: api.ToolFunction{Name: "test_tool"}},
+	}
+
+	lastMessage := &api.Message{Role: "assistant", Content: "previous"}
+
+	returnedTools := parser.Init(tools, lastMessage, nil)
+
+	if len(returnedTools) != len(tools) {
+		t.Errorf("expected %d tools returned, got %d", len(tools), len(returnedTools))
+	}
+}
+
+func TestCogitoParser_parseToolCallContent(t *testing.T) {
+	tests := []struct {
+		name        string
+		content     string
+		expected    api.ToolCall
+		expectError bool
+	}{
+		{
+			name: "valid_tool_call_standard_format",
+			content: `function<｜tool▁sep｜>get_weather
+` + "```json\n" + `{"location":"Paris"}
+` + "```",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "get_weather",
+					Arguments: api.ToolCallFunctionArguments{
+						"location": "Paris",
+					},
+				},
+			},
+			expectError: false,
+		},
+		{
+			name: "valid_tool_call_complex_args",
+			content: `function<｜tool▁sep｜>process_data
+` + "```json\n" + `{"items":["item1","item2"],"config":{"enabled":true},"count":42}
+` + "```",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "process_data",
+					Arguments: api.ToolCallFunctionArguments{
+						"items":  []any{"item1", "item2"},
+						"config": map[string]any{"enabled": true},
+						"count":  42.0,
+					},
+				},
+			},
+			expectError: false,
+		},
+		{
+			name: "valid_tool_call_empty_args",
+			content: `function<｜tool▁sep｜>no_args_tool
+` + "```json\n" + `{}
+` + "```",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      "no_args_tool",
+					Arguments: api.ToolCallFunctionArguments{},
+				},
+			},
+			expectError: false,
+		},
+		{
+			name:        "missing_separator",
+			content:     `functionget_weather` + "```json\n" + `{"location":"Paris"}` + "\n```",
+			expected:    api.ToolCall{},
+			expectError: true,
+		},
+		{
+			name:        "invalid_function_type",
+			content:     `not_function<｜tool▁sep｜>get_weather` + "```json\n" + `{"location":"Paris"}` + "\n```",
+			expected:    api.ToolCall{},
+			expectError: true,
+		},
+		{
+			name:        "missing_json_block_start",
+			content:     `function<｜tool▁sep｜>get_weather{"location":"Paris"}` + "```",
+			expected:    api.ToolCall{},
+			expectError: true,
+		},
+		{
+			name:        "missing_json_block_end",
+			content:     `function<｜tool▁sep｜>get_weather` + "```json\n" + `{"location":"Paris"}`,
+			expected:    api.ToolCall{},
+			expectError: true,
+		},
+		{
+			name:        "invalid_json",
+			content:     `function<｜tool▁sep｜>get_weather` + "```json\n" + `{location:Paris}` + "\n```",
+			expected:    api.ToolCall{},
+			expectError: true,
+		},
+		{
+			name:        "empty_function_type",
+			content:     `<｜tool▁sep｜>get_weather` + "```json\n" + `{"location":"Paris"}` + "\n```",
+			expected:    api.ToolCall{},
+			expectError: true,
+		},
+		{
+			name: "tool_with_spaces_in_name",
+			content: `function<｜tool▁sep｜>  get_weather  
+` + "```json\n" + `{"location":"Paris"}
+` + "```",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "get_weather",
+					Arguments: api.ToolCallFunctionArguments{
+						"location": "Paris",
+					},
+				},
+			},
+			expectError: false,
+		},
+		{
+			name: "tool_with_multiline_json",
+			content: `function<｜tool▁sep｜>get_weather
+` + "```json\n" + `{
+  "location": "Paris",
+  "units": "metric"
+}
+` + "```",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "get_weather",
+					Arguments: api.ToolCallFunctionArguments{
+						"location": "Paris",
+						"units":    "metric",
+					},
+				},
+			},
+			expectError: false,
+		},
+		{
+			name: "tool_with_nested_objects",
+			content: `function<｜tool▁sep｜>complex_tool
+` + "```json\n" + `{"nested":{"deep":{"value":123}}}
+` + "```",
+			expected: api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name: "complex_tool",
+					Arguments: api.ToolCallFunctionArguments{
+						"nested": map[string]any{
+							"deep": map[string]any{
+								"value": 123.0,
+							},
+						},
+					},
+				},
+			},
+			expectError: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			parser := &CogitoParser{}
+
+			result, err := parser.parseToolCallContent(tt.content)
+
+			if tt.expectError {
+				if err == nil {
+					t.Errorf("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("tool call mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}

model/parsers/parsers.go

@@ -6,9 +6,9 @@ import (
 )
 
 type Parser interface {
-	// Init initializes the parser with tools and optional last message for chat prefill
+	// Init initializes the parser with tools, optional last message for chat prefill, and think value
 	// Returns processed tools if the parser needs to modify them (e.g., harmony renames them)
-	Init(tools []api.Tool, lastMessage *api.Message) []api.Tool
+	Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool
 	// Add processes streamed content and returns parsed content, thinking, and tool calls
 	// The done flag indicates if this is the last chunk (used for draining accumulators)
 	Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error)
@@ -52,6 +52,8 @@ func ParserForName(name string) Parser {
 		return &PassthroughParser{}
 	case "harmony":
 		return harmony.NewHarmonyMessageHandler()
+	case "cogito":
+		return &CogitoParser{}
 	default:
 		return nil
 	}
@@ -59,7 +61,7 @@ func ParserForName(name string) Parser {
 
 type PassthroughParser struct{}
 
-func (p *PassthroughParser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
+func (p *PassthroughParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
 	return tools // passthrough doesn't modify tools
 }
 

model/parsers/parsers_test.go

@@ -10,7 +10,7 @@ type mockParser struct {
 	name string
 }
 
-func (m *mockParser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
+func (m *mockParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
 	return tools
 }
 

model/parsers/qwen3coder.go

@@ -43,7 +43,7 @@ func (p *Qwen3CoderParser) HasThinkingSupport() bool {
 	return false
 }
 
-func (p *Qwen3CoderParser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
+func (p *Qwen3CoderParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
 	p.tools = tools
 	return tools // Qwen doesn't modify tools
 }
@@ -432,7 +432,7 @@ func transformToXML(raw string) string {
 		groups := qwenTagRegex.FindStringSubmatch(match)
 		tag := groups[1]
 		var escapedValue strings.Builder
-		xml.EscapeText(&escapedValue, []byte(groups[2]))
+		_ = xml.EscapeText(&escapedValue, []byte(groups[2])) // error is always nil for strings.Builder
 		return fmt.Sprintf(`<%s name="%s">`, tag, escapedValue.String())
 	})
 

model/parsers/qwen3vl.go

@@ -54,7 +54,7 @@ func (p *Qwen3VLParser) setInitialState(lastMessage *api.Message) {
 	p.state = CollectingThinkingContent
 }
 
-func (p *Qwen3VLParser) Init(tools []api.Tool, lastMessage *api.Message) []api.Tool {
+func (p *Qwen3VLParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
 	p.tools = tools
 	p.setInitialState(lastMessage)
 	return tools

model/parsers/qwen3vl_nonthinking_test.go

@@ -198,7 +198,7 @@ func TestQwen3VLNonThinkingParserStreaming(t *testing.T) {
 
 		t.Run(tc.desc, func(t *testing.T) {
 			parser := Qwen3VLParser{hasThinkingSupport: false}
-			parser.Init([]api.Tool{}, nil)
+			parser.Init([]api.Tool{}, nil, nil)
 
 			for i, step := range tc.steps {
 				parser.buffer.WriteString(step.input)
@@ -515,7 +515,7 @@ func TestQwenOldParserStreaming(t *testing.T) {
 
 		t.Run(tc.desc, func(t *testing.T) {
 			parser := Qwen3VLParser{hasThinkingSupport: false}
-			parser.Init([]api.Tool{}, nil)
+			parser.Init([]api.Tool{}, nil, nil)
 
 			for i, step := range tc.steps {
 				parser.buffer.WriteString(step.input)
@@ -822,7 +822,7 @@ func TestQwen3VLNonThinkingToolCallWhitespaceHandling(t *testing.T) {
 
 		t.Run(tc.desc, func(t *testing.T) {
 			parser := Qwen3VLParser{hasThinkingSupport: false}
-			parser.Init([]api.Tool{}, nil)
+			parser.Init([]api.Tool{}, nil, nil)
 
 			for i, step := range tc.steps {
 				parser.buffer.WriteString(step.input)

model/parsers/qwen3vl_thinking_test.go

@@ -205,7 +205,7 @@ func TestQwen3VLThinkingParserStreaming(t *testing.T) {
 
 		t.Run(tc.desc, func(t *testing.T) {
 			parser := Qwen3VLParser{hasThinkingSupport: true}
-			parser.Init([]api.Tool{}, nil)
+			parser.Init([]api.Tool{}, nil, nil)
 			// parser.state = CollectingThinkingContent
 
 			for i, step := range tc.steps {
@@ -386,7 +386,7 @@ func TestQwen3VLParserState(t *testing.T) {
 
 	for _, tc := range cases {
 		parser := Qwen3VLParser{hasThinkingSupport: tc.hasThinking}
-		parser.Init(nil, tc.last)
+		parser.Init(nil, tc.last, nil)
 		if parser.state != tc.wantState {
 			t.Errorf("%s: got state %v, want %v", tc.desc, parser.state, tc.wantState)
 		}
@@ -437,7 +437,7 @@ func TestQwen3VLThinkingParserWithThinkingPrefill(t *testing.T) {
 	for _, tc := range cases {
 		t.Run(tc.desc, func(t *testing.T) {
 			parser := Qwen3VLParser{hasThinkingSupport: true}
-			parser.Init([]api.Tool{}, last)
+			parser.Init([]api.Tool{}, last, nil)
 
 			for i, step := range tc.steps {
 				parser.buffer.WriteString(step.input)
@@ -500,7 +500,7 @@ func TestQwen3VLThinkingParserWithNonThinkingPrefill(t *testing.T) {
 	for _, tc := range cases {
 		t.Run(tc.desc, func(t *testing.T) {
 			parser := Qwen3VLParser{hasThinkingSupport: true}
-			parser.Init([]api.Tool{}, last)
+			parser.Init([]api.Tool{}, last, nil)
 
 			for i, step := range tc.steps {
 				parser.buffer.WriteString(step.input)
@@ -523,7 +523,7 @@ func TestQwen3VLThinkingParserStreamingAssistantPrefillContent(t *testing.T) {
 	// last message is assistant with content ⇒ start in CollectingContent
 	last := &api.Message{Role: "assistant", Content: "has content"}
 	parser := Qwen3VLParser{hasThinkingSupport: true}
-	parser.Init([]api.Tool{}, last)
+	parser.Init([]api.Tool{}, last, nil)
 
 	type step struct {
 		input      string
@@ -750,7 +750,7 @@ func TestQwen3VLThinkingWhitespaceHandling(t *testing.T) {
 
 		t.Run(tc.desc, func(t *testing.T) {
 			parser := Qwen3VLParser{hasThinkingSupport: true}
-			parser.Init([]api.Tool{}, nil)
+			parser.Init([]api.Tool{}, nil, nil)
 
 			for i, step := range tc.steps {
 				parser.buffer.WriteString(step.input)
@@ -859,7 +859,7 @@ func TestQwen3VLToolCallWhitespaceHandling(t *testing.T) {
 
 		t.Run(tc.desc, func(t *testing.T) {
 			parser := Qwen3VLParser{hasThinkingSupport: true}
-			parser.Init([]api.Tool{}, tc.prefillMsg)
+			parser.Init([]api.Tool{}, tc.prefillMsg, nil)
 
 			for i, step := range tc.steps {
 				parser.buffer.WriteString(step.input)

server/routes.go

@@ -340,7 +340,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		builtinParser = parsers.ParserForName(m.Config.Parser)
 		if builtinParser != nil {
 			// no tools or last message for generate endpoint
-			builtinParser.Init(nil, nil)
+			builtinParser.Init(nil, nil, req.Think)
 		}
 	}
 
@@ -2051,7 +2051,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 				lastMessage = &msgs[len(msgs)-1]
 			}
 			// Initialize parser and get processed tools
-			processedTools = builtinParser.Init(req.Tools, lastMessage)
+			processedTools = builtinParser.Init(req.Tools, lastMessage, req.Think)
 		}
 	}