renderers: add global flag for setting [img] tags (#12669)

Adds a temporary global flag to renderers that causes renderers to always
render images as [img]. In a follow up change, we will consider making this
the default, and this flag could eventually be removed

model/renderers/qwen3vl.go

@@ -48,13 +48,22 @@ func marshalWithSpaces(v any) ([]byte, error) {
 
 type Qwen3VLRenderer struct {
 	isThinking bool
+
+	useImgTags bool
 }
 
-func (r *Qwen3VLRenderer) renderContent(content api.Message, doVisionCount bool) string {
+func (r *Qwen3VLRenderer) renderContent(content api.Message) string {
 	// This assumes all images are at the front of the message - same assumption as ollama/ollama/runner.go
 	var subSb strings.Builder
 	for range content.Images {
-		subSb.WriteString("<|vision_start|><|image_pad|><|vision_end|>")
+		// TODO: (jmorganca): how to render this is different for different
+		// model backends, and so we should eventually parameterize this or
+		// only output a placeholder such as [img]
+		if r.useImgTags {
+			subSb.WriteString("[img]")
+		} else {
+			subSb.WriteString("<|vision_start|><|image_pad|><|vision_end|>")
+		}
 	}
 	// TODO: support videos
 
@@ -88,7 +97,7 @@ func (r *Qwen3VLRenderer) Render(messages []api.Message, tools []api.Tool, _ *ap
 		message := messages[i]
 		if multiStepTool && message.Role == "user" {
 			// Check if content starts with <tool_response> and ends with </tool_response>
-			content := r.renderContent(message, true)
+			content := r.renderContent(message)
 			if !(strings.HasPrefix(content, "<tool_response>") && strings.HasSuffix(content, "</tool_response>")) {
 				multiStepTool = false
 				lastQueryIndex = i
@@ -97,7 +106,7 @@ func (r *Qwen3VLRenderer) Render(messages []api.Message, tools []api.Tool, _ *ap
 	}
 
 	for i, message := range messages {
-		content := r.renderContent(message, true)
+		content := r.renderContent(message)
 
 		lastMessage := i == len(messages)-1
 		prefill := lastMessage && message.Role == "assistant"

model/renderers/qwen3vl_nonthinking_test.go

@@ -9,11 +9,12 @@ import (
 
 func TestQwen3VLNonThinkingRenderer(t *testing.T) {
 	tests := []struct {
-		name     string
-		msgs     []api.Message
-		images   []api.ImageData
-		tools    []api.Tool
-		expected string
+		name       string
+		msgs       []api.Message
+		images     []api.ImageData
+		tools      []api.Tool
+		useImgTags bool
+		expected   string
 	}{
 		{
 			name: "prefill",
@@ -90,6 +91,18 @@ I'll check the weather in San Francisco for you.<think>Speak poetry after the fi
 			expected: `<|im_start|>user
 <|vision_start|><|image_pad|><|vision_end|>Describe this image.<|im_end|>
 <|im_start|>assistant
+Let me analyze this image.`,
+		},
+		{
+			name: "Image with image tags",
+			msgs: []api.Message{
+				{Role: "user", Content: "Describe this image.", Images: []api.ImageData{api.ImageData("img2")}},
+				{Role: "assistant", Content: "Let me analyze this image."},
+			},
+			useImgTags: true,
+			expected: `<|im_start|>user
+[img]Describe this image.<|im_end|>
+<|im_start|>assistant
 Let me analyze this image.`,
 		},
 		{
@@ -102,7 +115,18 @@ Let me analyze this image.`,
 <|im_start|>assistant
 `,
 		},
-
+		{
+			name: "Multiple images with image tags",
+			msgs: []api.Message{
+				{Role: "user", Content: "Describe these images.", Images: []api.ImageData{api.ImageData("img1"), api.ImageData("img2")}},
+				{Role: "assistant", Content: "Let me analyze this image."},
+			},
+			useImgTags: true,
+			expected: `<|im_start|>user
+[img][img]Describe these images.<|im_end|>
+<|im_start|>assistant
+Let me analyze this image.`,
+		},
 		// 		// NOTE: solved with #12518: https://github.com/ollama/ollama/compare/main...drifkin/stable-tool-args
 		// 		{
 		// 			name: "with tools and response",
@@ -485,7 +509,7 @@ I'll check.
 	}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			rendered, err := (&Qwen3VLRenderer{false}).Render(tt.msgs, tt.tools, nil)
+			rendered, err := (&Qwen3VLRenderer{isThinking: false, useImgTags: tt.useImgTags}).Render(tt.msgs, tt.tools, nil)
 			if err != nil {
 				t.Fatal(err)
 			}

model/renderers/qwen3vl_thinking_test.go

@@ -323,7 +323,7 @@ Speak poetry after the first sentence.</think><think>Speak poetry after the seco
 	}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			rendered, err := (&Qwen3VLRenderer{true}).Render(tt.msgs, tt.tools, nil)
+			rendered, err := (&Qwen3VLRenderer{isThinking: true}).Render(tt.msgs, tt.tools, nil)
 			if err != nil {
 				t.Fatal(err)
 			}

model/renderers/renderer.go

@@ -17,6 +17,11 @@ type (
 	}
 )
 
+// RenderImgTags is a global flag that tells renderers to use [img] tags
+// for images. This is set by the Ollama server package on init, or left as
+// false for other environments where renderers are used
+var RenderImgTags bool
+
 func (r *RendererRegistry) Register(name string, renderer RendererConstructor) {
 	r.renderers[name] = renderer
 }
@@ -46,10 +51,10 @@ func rendererForName(name string) Renderer {
 		renderer := &Qwen3CoderRenderer{}
 		return renderer
 	case "qwen3-vl-instruct":
-		renderer := &Qwen3VLRenderer{false}
+		renderer := &Qwen3VLRenderer{isThinking: false, useImgTags: RenderImgTags}
 		return renderer
 	case "qwen3-vl-thinking":
-		renderer := &Qwen3VLRenderer{true}
+		renderer := &Qwen3VLRenderer{isThinking: true, useImgTags: RenderImgTags}
 		return renderer
 	default:
 		return nil

server/routes.go

@@ -39,6 +39,7 @@ import (
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/middleware"
 	"github.com/ollama/ollama/model/parsers"
+	"github.com/ollama/ollama/model/renderers"
 	"github.com/ollama/ollama/server/internal/client/ollama"
 	"github.com/ollama/ollama/server/internal/registry"
 	"github.com/ollama/ollama/template"
@@ -91,6 +92,9 @@ func init() {
 	}
 
 	gin.SetMode(mode)
+
+	// Tell renderers to use [img] tags
+	renderers.RenderImgTags = true
 }
 
 var (