server: align generate with native chat templates (#16878)

* server: align generate with native chat templates

/api/generate rebuilt chat-like prompts through the Go template path even when the model selected its native GGUF Jinja chat template, so the same model rendered differently between generate and chat.

Route chat-like generate requests through the shared native chat preparation path, keep deprecated context and image handling working there, and keep explicit OLLAMA_GO_TEMPLATE overrides intact.

Fixes #16792

* review comments

Fall back to "{{ .Prompt }}" when lacking templates

integration/generate_jinja_test.go

@@ -0,0 +1,260 @@
+//go:build integration && generate
+
+package integration
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestGenerateNativeJinjaDebugRender(t *testing.T) {
+	if os.Getenv("OLLAMA_TEST_EXISTING") != "" {
+		t.Skip("requires a test-managed server started with OLLAMA_GO_TEMPLATE=0")
+	}
+	t.Setenv("OLLAMA_GO_TEMPLATE", "0")
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
+	defer cancel()
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	pullOrSkip(ctx, t, client, smol)
+
+	modelName := createGenerateJinjaModel(ctx, t, client, smol)
+	defer func() {
+		if err := client.Delete(context.Background(), &api.DeleteRequest{Model: modelName}); err != nil {
+			t.Logf("failed to delete %s: %v", modelName, err)
+		}
+	}()
+
+	t.Run("generate matches chat", func(t *testing.T) {
+		prompt := "Reply with one short sentence about blue skies."
+		generateRender := debugGenerateRender(ctx, t, client, api.GenerateRequest{
+			Model:  modelName,
+			Prompt: prompt,
+		})
+		chatRender := debugChatRender(ctx, t, client, api.ChatRequest{
+			Model: modelName,
+			Messages: []api.Message{
+				{Role: "user", Content: prompt},
+			},
+		})
+
+		if generateRender == "" {
+			t.Fatal("generate debug render was empty")
+		}
+		if strings.Contains(generateRender, "GO_TEMPLATE_SENTINEL") {
+			t.Fatalf("generate used Go TEMPLATE instead of native chat template: %q", generateRender)
+		}
+		if generateRender != chatRender {
+			t.Fatalf("generate render does not match chat render\ngenerate: %q\nchat:     %q", generateRender, chatRender)
+		}
+	})
+
+	t.Run("context prepends prior tokens", func(t *testing.T) {
+		first := generateOnce(ctx, t, client, api.GenerateRequest{
+			Model:  modelName,
+			Prompt: "Say hello.",
+			Options: map[string]any{
+				"num_predict": 1,
+				"temperature": 0,
+			},
+		})
+		if len(first.Context) == 0 {
+			t.Fatal("generate response did not include context")
+		}
+
+		next := api.GenerateRequest{
+			Model:  modelName,
+			Prompt: "Say goodbye.",
+		}
+		withoutContext := debugGenerateRender(ctx, t, client, next)
+		next.Context = first.Context
+		withContext := debugGenerateRender(ctx, t, client, next)
+
+		if withContext == withoutContext {
+			t.Fatal("context debug render did not change the prompt")
+		}
+		if !strings.HasSuffix(withContext, withoutContext) {
+			t.Fatalf("context render should preserve current native template as suffix\nwith context: %q\nwithout:      %q", withContext, withoutContext)
+		}
+	})
+
+	t.Run("images use completion media markers", func(t *testing.T) {
+		resp := debugGenerate(ctx, t, client, api.GenerateRequest{
+			Model:  modelName,
+			Prompt: "compare [img] and [img]",
+			Images: []api.ImageData{
+				[]byte("first-test-image"),
+				[]byte("second-test-image"),
+			},
+		})
+
+		if resp.DebugInfo == nil {
+			t.Fatal("missing debug info")
+		}
+		if resp.DebugInfo.ImageCount != 2 {
+			t.Fatalf("image count = %d, want 2", resp.DebugInfo.ImageCount)
+		}
+		rendered := resp.DebugInfo.RenderedTemplate
+		if !strings.Contains(rendered, "[img-0]") || !strings.Contains(rendered, "[img-1]") {
+			t.Fatalf("rendered template missing image markers: %q", rendered)
+		}
+	})
+}
+
+func TestGenerateNativeJinjaImages(t *testing.T) {
+	if os.Getenv("OLLAMA_TEST_EXISTING") != "" {
+		t.Skip("requires a test-managed server started with OLLAMA_GO_TEMPLATE=0")
+	}
+	t.Setenv("OLLAMA_GO_TEMPLATE", "0")
+	skipUnderMinVRAM(t, 6)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
+	defer cancel()
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	_, _, ollamaHome := decodeTestImages(t)
+
+	t.Run("actual image processing", func(t *testing.T) {
+		for _, baseModel := range testModels([]string{"gemma3:4b"}) {
+			t.Run(baseModel, func(t *testing.T) {
+				setupVisionModel(ctx, t, client, baseModel)
+
+				modelName := createGenerateJinjaModel(ctx, t, client, baseModel)
+				defer func() {
+					if err := client.Delete(context.Background(), &api.DeleteRequest{Model: modelName}); err != nil {
+						t.Logf("failed to delete %s: %v", modelName, err)
+					}
+				}()
+
+				req := api.GenerateRequest{
+					Model:  modelName,
+					Prompt: "Describe what you see in this image briefly.",
+					Images: []api.ImageData{ollamaHome},
+					Options: map[string]any{
+						"seed":        42,
+						"temperature": 0.0,
+					},
+					KeepAlive: &api.Duration{Duration: 10 * time.Second},
+				}
+
+				resp := debugGenerate(ctx, t, client, req)
+				if resp.DebugInfo == nil {
+					t.Fatal("missing debug info")
+				}
+				if resp.DebugInfo.ImageCount != 1 {
+					t.Fatalf("image count = %d, want 1", resp.DebugInfo.ImageCount)
+				}
+				rendered := resp.DebugInfo.RenderedTemplate
+				if strings.Contains(rendered, "GO_TEMPLATE_SENTINEL") {
+					t.Fatalf("generate with images used Go TEMPLATE instead of native chat template: %q", rendered)
+				}
+				if !strings.Contains(rendered, "[img-0]") {
+					t.Fatalf("rendered template missing image marker: %q", rendered)
+				}
+
+				DoGenerate(ctx, t, client, req, []string{
+					"llama", "animal", "build", "model", "open", "cartoon", "character",
+				}, 120*time.Second, 30*time.Second)
+			})
+		}
+	})
+}
+
+func createGenerateJinjaModel(ctx context.Context, t *testing.T, client *api.Client, baseModel string) string {
+	t.Helper()
+
+	modelName := fmt.Sprintf("test-generate-jinja-%d", time.Now().UnixNano())
+	stream := false
+	err := client.Create(ctx, &api.CreateRequest{
+		Model:    modelName,
+		From:     baseModel,
+		Stream:   &stream,
+		Template: `GO_TEMPLATE_SENTINEL {{ if .System }}{{ .System }} {{ end }}{{ if .Prompt }}{{ .Prompt }}{{ end }}{{ range .Messages }}{{ .Content }}{{ end }}`,
+	}, func(api.ProgressResponse) error {
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("create %s from %s: %v", modelName, baseModel, err)
+	}
+	return modelName
+}
+
+func debugGenerateRender(ctx context.Context, t *testing.T, client *api.Client, req api.GenerateRequest) string {
+	t.Helper()
+	resp := debugGenerate(ctx, t, client, req)
+	if resp.DebugInfo == nil {
+		t.Fatal("missing debug info")
+	}
+	return resp.DebugInfo.RenderedTemplate
+}
+
+func debugGenerate(ctx context.Context, t *testing.T, client *api.Client, req api.GenerateRequest) api.GenerateResponse {
+	t.Helper()
+
+	stream := false
+	req.Stream = &stream
+	req.DebugRenderOnly = true
+
+	var final api.GenerateResponse
+	if err := client.Generate(ctx, &req, func(resp api.GenerateResponse) error {
+		final = resp
+		return nil
+	}); err != nil {
+		if strings.Contains(err.Error(), "requires more system memory") {
+			t.Skipf("model is too large for this test system: %v", err)
+		}
+		t.Fatalf("generate debug render failed: %v", err)
+	}
+	return final
+}
+
+func debugChatRender(ctx context.Context, t *testing.T, client *api.Client, req api.ChatRequest) string {
+	t.Helper()
+
+	stream := false
+	req.Stream = &stream
+	req.DebugRenderOnly = true
+
+	var final api.ChatResponse
+	if err := client.Chat(ctx, &req, func(resp api.ChatResponse) error {
+		final = resp
+		return nil
+	}); err != nil {
+		if strings.Contains(err.Error(), "requires more system memory") {
+			t.Skipf("model is too large for this test system: %v", err)
+		}
+		t.Fatalf("chat debug render failed: %v", err)
+	}
+	if final.DebugInfo == nil {
+		t.Fatal("missing debug info")
+	}
+	return final.DebugInfo.RenderedTemplate
+}
+
+func generateOnce(ctx context.Context, t *testing.T, client *api.Client, req api.GenerateRequest) api.GenerateResponse {
+	t.Helper()
+
+	stream := false
+	req.Stream = &stream
+
+	var final api.GenerateResponse
+	if err := client.Generate(ctx, &req, func(resp api.GenerateResponse) error {
+		final = resp
+		return nil
+	}); err != nil {
+		if strings.Contains(err.Error(), "requires more system memory") {
+			t.Skipf("model is too large for this test system: %v", err)
+		}
+		t.Fatalf("generate failed: %v", err)
+	}
+	return final
+}

server/images.go

@@ -272,13 +272,32 @@ func hasMoreCapabilities(candidate, current []model.Capability) bool {
 	return len(candidate) > len(current)
 }
 
-func shouldPreferChatTemplate(chatTemplate string, chatTemplateCaps []model.Capability, goTemplate *template.Template, goTemplateCaps []model.Capability) bool {
-	if !hasMoreCapabilities(chatTemplateCaps, goTemplateCaps) {
+func sameCapabilities(candidate, current []model.Capability) bool {
+	if len(candidate) != len(current) {
 		return false
 	}
+	for _, c := range candidate {
+		if !slices.Contains(current, c) {
+			return false
+		}
+	}
+	return true
+}
+
+func shouldPreferChatTemplate(chatTemplate string, chatTemplateCaps []model.Capability, goTemplate *template.Template, goTemplateCaps []model.Capability) bool {
+	if hasMoreCapabilities(chatTemplateCaps, goTemplateCaps) {
 		return !goTemplateHasToolRoundTrip(goTemplate) || chatTemplateHasToolRoundTrip(chatTemplate)
 	}
 
+	if !sameCapabilities(chatTemplateCaps, goTemplateCaps) ||
+		!slices.Contains(chatTemplateCaps, model.CapabilityTools) ||
+		!slices.Contains(goTemplateCaps, model.CapabilityTools) {
+		return false
+	}
+
+	return chatTemplateHasToolRoundTrip(chatTemplate) && !goTemplateHasToolRoundTrip(goTemplate)
+}
+
 func goTemplateEnvSet() bool {
 	return envconfig.GoTemplate(true) == envconfig.GoTemplate(false)
 }

server/images_test.go

@@ -132,6 +132,37 @@ func TestGetModelTemplateMetadata(t *testing.T) {
 		}
 	})
 
+	t.Run("prefers chat template with stronger tool round trip", func(t *testing.T) {
+		t.Setenv("OLLAMA_MODELS", t.TempDir())
+		t.Setenv("OLLAMA_GO_TEMPLATE", "")
+
+		_, digest := createBinFile(t, ggml.KV{
+			"general.architecture": "llama",
+			"tokenizer.chat_template": `{% if tools %}{{ tools }}{% endif %}
+{% for message in messages %}
+{% if message.tool_calls %}
+{% for tool_call in message.tool_calls %}{{ tool_call.function.name }}{% endfor %}
+{% endif %}
+{% if message.role == 'tool' %}tool_response {{ message.content }}{% endif %}
+{% endfor %}`,
+		}, nil)
+		writeTestModelManifest(t, "chat-template-tool-round-trip", digest, `{{ if .Tools }}tools{{ end }}
+{{ range .Messages }}
+{{ range .ToolCalls }}{{ .Function.Name }}{{ end }}
+{{ end }}`)
+
+		m, err := GetModel("chat-template-tool-round-trip")
+		if err != nil {
+			t.Fatal(err)
+		}
+		if !m.PreferChatTemplate {
+			t.Fatal("expected chat template to be preferred")
+		}
+		if got := m.CheckCapabilities(model.CapabilityTools); got != nil {
+			t.Fatalf("expected tools capability, got %v", got)
+		}
+	})
+
 	t.Run("keeps Go TEMPLATE when chat template has weaker tool support", func(t *testing.T) {
 		t.Setenv("OLLAMA_MODELS", t.TempDir())
 		t.Setenv("OLLAMA_GO_TEMPLATE", "")

server/prompt.go

@@ -77,11 +77,27 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 		slog.Debug("truncating input messages which exceed context length", "truncated", len(msgs[currMsgIdx:]))
 	}
 
-	renderMsgs := slices.Clone(msgs)
+	renderMsgs, media, err := imageTaggedMessages(m, msgs, currMsgIdx, false)
+	if err != nil {
+		return "", nil, err
+	}
 
-	for cnt, msg := range renderMsgs[currMsgIdx:] {
+	// truncate any messages that do not fit into the context window
+	p, err := renderPrompt(m, append(system, renderMsgs[currMsgIdx:]...), tools, think)
+	if err != nil {
+		return "", nil, err
+	}
+
+	return p, media, nil
+}
+
+func imageTaggedMessages(m *Model, msgs []api.Message, start int, clearImages bool) ([]api.Message, []llm.MediaData, error) {
+	renderMsgs := slices.Clone(msgs)
+	var media []llm.MediaData
+
+	for cnt, msg := range renderMsgs[start:] {
 		if slices.Contains(m.Config.ModelFamilies, "mllama") && len(msg.Images) > 1 {
-			return "", nil, errors.New("this model only supports one image while more than one image requested")
+			return nil, nil, errors.New("this model only supports one image while more than one image requested")
 		}
 
 		var prefix string
@@ -105,20 +121,15 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 			}
 		}
 
-		if m.Config.Renderer != "" {
-			continue
+		if m.Config.Renderer == "" {
+			renderMsgs[start+cnt].Content = prefix + prompt
+		}
+		if clearImages {
+			renderMsgs[start+cnt].Images = nil
+		}
 	}
 
-		renderMsgs[currMsgIdx+cnt].Content = prefix + prompt
-	}
-
-	// truncate any messages that do not fit into the context window
-	p, err := renderPrompt(m, append(system, renderMsgs[currMsgIdx:]...), tools, think)
-	if err != nil {
-		return "", nil, err
-	}
-
-	return p, media, nil
+	return renderMsgs, media, nil
 }
 
 func renderPrompt(m *Model, msgs []api.Message, tools []api.Tool, think *api.ThinkValue) (string, error) {

server/routes.go

@@ -567,6 +567,47 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		// support for generate
 		if values.Messages != nil && values.Suffix == "" && req.Template == "" {
 			genTruncate := (req.Truncate == nil || *req.Truncate) && !m.IsMLX()
+			if m.HasChatTemplate && chatModeForModel(m) == chatExecutionModeNative {
+				nativeReq, err := prepareNativeChatRequest(c.Request.Context(), m, r, opts, llm.ChatRequest{
+					Messages:    values.Messages,
+					Format:      req.Format,
+					Options:     opts,
+					Think:       req.Think,
+					Shift:       req.Shift == nil || *req.Shift,
+					Logprobs:    req.Logprobs,
+					TopLogprobs: req.TopLogprobs,
+				}, genTruncate)
+				if err != nil {
+					slog.Error("chat template prompt error", "error", err)
+					var serr api.StatusError
+					if errors.As(err, &serr) {
+						c.JSON(serr.StatusCode, gin.H{"error": serr.ErrorMessage})
+					} else {
+						c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+					}
+					return
+				}
+				nativeReq.Messages, media, err = imageTaggedMessages(m, nativeReq.Messages, 0, true)
+				if err != nil {
+					c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+					return
+				}
+				prompt, err = r.ApplyChatTemplate(c.Request.Context(), nativeReq)
+				if err != nil {
+					var serr api.StatusError
+					if errors.As(err, &serr) {
+						c.JSON(serr.StatusCode, gin.H{"error": serr.ErrorMessage})
+					} else {
+						c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+					}
+					return
+				}
+				// TEMP(drifkin): req.Context will be removed very soon, but we're temporarily supporting it in this flow here
+				if req.Context != nil {
+					b.WriteString(prompt)
+					prompt = b.String()
+				}
+			} else {
 				prompt, media, err = chatPrompt(c.Request.Context(), m, r.Tokenize, optionsForPrompt(opts, r), values.Messages, []api.Tool{}, req.Think, genTruncate)
 				if err != nil {
 					c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
@@ -578,6 +619,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 					prompt = b.String()
 				}
 				leadingBOS = leadingBOSForModel(m)
+			}
 		} else {
 			// Direct template execution flow.
 			if err := tmpl.Execute(&b, values); err != nil {
@@ -2917,8 +2959,15 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	writeChatResponse(c, req, ch)
 }
 
+func prepareNativeChatRequest(ctx context.Context, m *Model, r llm.LlamaServer, opts *api.Options, nativeReq llm.ChatRequest, truncate bool) (llm.ChatRequest, error) {
+	var err error
+	nativeReq.Messages, err = truncateNativeChatMessages(ctx, m, r, optionsForPrompt(opts, r), nativeReq, truncate)
+	return nativeReq, err
+}
+
 func (s *Server) handleNativeChat(c *gin.Context, req api.ChatRequest, m *Model, r llm.LlamaServer, opts *api.Options, msgs []api.Message, checkpointStart, checkpointLoaded time.Time) {
-	nativeReq := llm.ChatRequest{
+	truncate := req.Truncate == nil || *req.Truncate
+	nativeReq, err := prepareNativeChatRequest(c.Request.Context(), m, r, opts, llm.ChatRequest{
 		Messages:    msgs,
 		Tools:       req.Tools,
 		Format:      req.Format,
@@ -2927,10 +2976,7 @@ func (s *Server) handleNativeChat(c *gin.Context, req api.ChatRequest, m *Model,
 		Shift:       req.Shift == nil || *req.Shift,
 		Logprobs:    req.Logprobs,
 		TopLogprobs: req.TopLogprobs,
-	}
-	truncate := req.Truncate == nil || *req.Truncate
-	var err error
-	nativeReq.Messages, err = truncateNativeChatMessages(c.Request.Context(), m, r, optionsForPrompt(opts, r), nativeReq, truncate)
+	}, truncate)
 	if err != nil {
 		slog.Error("chat template prompt error", "error", err)
 		var serr api.StatusError

server/routes_generate_test.go

@@ -62,6 +62,7 @@ type mockRunner struct {
 	ChatFn        func(context.Context, llm.ChatRequest, func(llm.ChatResponse)) error
 	Template      string
 	TemplateFn    func(context.Context, llm.ChatRequest) (string, error)
+	DetokenizeFn  func(context.Context, []int) (string, error)
 	contextLength int
 }
 
@@ -91,6 +92,13 @@ func (m *mockRunner) ApplyChatTemplate(ctx context.Context, r llm.ChatRequest) (
 	return m.Template, nil
 }
 
+func (m *mockRunner) Detokenize(ctx context.Context, tokens []int) (string, error) {
+	if m.DetokenizeFn != nil {
+		return m.DetokenizeFn(ctx, tokens)
+	}
+	return "", nil
+}
+
 func (mockRunner) Tokenize(_ context.Context, s string) (tokens []int, err error) {
 	for range strings.Fields(s) {
 		tokens = append(tokens, len(tokens))
@@ -465,6 +473,200 @@ func TestChatHandlerHarmonyPreservesStructuralTokens(t *testing.T) {
 	}
 }
 
+func TestGenerateHandlerChatTemplateRoute(t *testing.T) {
+	t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
+	t.Setenv("OLLAMA_GO_TEMPLATE", "")
+	gin.SetMode(gin.TestMode)
+
+	t.Run("uses GGUF chat_template when no Go TEMPLATE exists", func(t *testing.T) {
+		mock := mockRunner{
+			TemplateFn: func(_ context.Context, req llm.ChatRequest) (string, error) {
+				if len(req.Messages) != 1 || req.Messages[0].Role != "user" || req.Messages[0].Content != "hello" {
+					t.Fatalf("chat template messages = %#v", req.Messages)
+				}
+				return "native-template: hello", nil
+			},
+			CompletionResponse: llm.CompletionResponse{
+				Content:            "ok",
+				Done:               true,
+				DoneReason:         llm.DoneReasonStop,
+				PromptEvalCount:    1,
+				PromptEvalDuration: time.Millisecond,
+				EvalCount:          1,
+				EvalDuration:       time.Millisecond,
+			},
+		}
+		s := newServerWithMockRunner(t, &mock)
+		createMinimalGGUFModel(t, s, "generate-chat-template", ggml.KV{
+			"tokenizer.chat_template": "{{ messages[0]['content'] }}",
+		}, "", nil)
+
+		stream := false
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "generate-chat-template",
+			Prompt: "hello",
+			Stream: &stream,
+		})
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+		}
+		if mock.CompletionRequest.Prompt != "native-template: hello" {
+			t.Fatalf("completion prompt = %q, want native chat_template prompt", mock.CompletionRequest.Prompt)
+		}
+		if mock.CompletionRequest.LeadingBOS != "" {
+			t.Fatalf("native chat_template prompt should not add leading BOS, got %q", mock.CompletionRequest.LeadingBOS)
+		}
+	})
+
+	t.Run("uses preferred GGUF chat_template over Go TEMPLATE", func(t *testing.T) {
+		mock := mockRunner{
+			TemplateFn: func(_ context.Context, req llm.ChatRequest) (string, error) {
+				if len(req.Messages) != 1 || req.Messages[0].Content != "hello" {
+					t.Fatalf("chat template messages = %#v", req.Messages)
+				}
+				return "preferred-native-template: hello", nil
+			},
+		}
+		s := newServerWithMockRunner(t, &mock)
+		createMinimalGGUFModel(t, s, "generate-preferred-chat-template", ggml.KV{
+			"tokenizer.chat_template": "{% if tools %}{{ tools }}{% endif %}{{ messages[0]['content'] }}",
+		}, "{{ range .Messages }}go-template: {{ .Content }}{{ end }}", nil)
+
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:           "generate-preferred-chat-template",
+			Prompt:          "hello",
+			DebugRenderOnly: true,
+		})
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+		}
+
+		var actual api.GenerateResponse
+		if err := json.Unmarshal(w.Body.Bytes(), &actual); err != nil {
+			t.Fatal(err)
+		}
+		if actual.DebugInfo == nil || actual.DebugInfo.RenderedTemplate != "preferred-native-template: hello" {
+			t.Fatalf("rendered template = %#v, want preferred native chat_template", actual.DebugInfo)
+		}
+	})
+
+	t.Run("falls back to default raw-like template without chat_template", func(t *testing.T) {
+		mock := mockRunner{
+			TemplateFn: func(context.Context, llm.ChatRequest) (string, error) {
+				t.Fatal("native chat template should not be used without tokenizer.chat_template")
+				return "", nil
+			},
+		}
+		s := newServerWithMockRunner(t, &mock)
+		createMinimalGGUFModel(t, s, "generate-no-chat-template", nil, "", nil)
+
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:           "generate-no-chat-template",
+			Prompt:          "plain prompt",
+			DebugRenderOnly: true,
+		})
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+		}
+
+		var actual api.GenerateResponse
+		if err := json.Unmarshal(w.Body.Bytes(), &actual); err != nil {
+			t.Fatal(err)
+		}
+		if actual.DebugInfo == nil || actual.DebugInfo.RenderedTemplate != "plain prompt" {
+			t.Fatalf("rendered template = %#v, want raw-like default template", actual.DebugInfo)
+		}
+	})
+
+	t.Run("prepends deprecated context to native chat template prompt", func(t *testing.T) {
+		mock := mockRunner{
+			TemplateFn: func(_ context.Context, req llm.ChatRequest) (string, error) {
+				if len(req.Messages) != 1 || req.Messages[0].Content != "next" {
+					t.Fatalf("chat template messages = %#v", req.Messages)
+				}
+				return "native-template: next", nil
+			},
+			DetokenizeFn: func(_ context.Context, tokens []int) (string, error) {
+				if !slices.Equal(tokens, []int{1, 2}) {
+					t.Fatalf("detokenize tokens = %#v", tokens)
+				}
+				return "prior context ", nil
+			},
+			CompletionResponse: llm.CompletionResponse{
+				Content:    "ok",
+				Done:       true,
+				DoneReason: llm.DoneReasonStop,
+			},
+		}
+		s := newServerWithMockRunner(t, &mock)
+		createMinimalGGUFModel(t, s, "generate-chat-template-context", ggml.KV{
+			"tokenizer.chat_template": "{{ messages[0]['content'] }}",
+		}, "", nil)
+
+		stream := false
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:   "generate-chat-template-context",
+			Prompt:  "next",
+			Context: []int{1, 2},
+			Stream:  &stream,
+		})
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+		}
+		if mock.CompletionRequest.Prompt != "prior context native-template: next" {
+			t.Fatalf("completion prompt = %q, want context-prefixed native chat_template prompt", mock.CompletionRequest.Prompt)
+		}
+	})
+
+	t.Run("keeps generate media markers with native chat template", func(t *testing.T) {
+		image1 := []byte("image-one")
+		image2 := []byte("image-two")
+		mock := mockRunner{
+			TemplateFn: func(_ context.Context, req llm.ChatRequest) (string, error) {
+				if len(req.Messages) != 1 {
+					t.Fatalf("chat template messages = %#v", req.Messages)
+				}
+				if len(req.Messages[0].Images) != 0 {
+					t.Fatalf("native generate render request should carry image markers, not image payloads: %#v", req.Messages[0])
+				}
+				return req.Messages[0].Content, nil
+			},
+			CompletionResponse: llm.CompletionResponse{
+				Content:    "ok",
+				Done:       true,
+				DoneReason: llm.DoneReasonStop,
+			},
+		}
+		s := newServerWithMockRunner(t, &mock)
+		createMinimalGGUFModel(t, s, "generate-chat-template-images", ggml.KV{
+			"tokenizer.chat_template": "{{ messages[0]['content'] }}",
+		}, "", nil)
+
+		stream := false
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "generate-chat-template-images",
+			Prompt: "compare [img] and [img]",
+			Images: []api.ImageData{image1, image2},
+			Stream: &stream,
+		})
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d: %s", w.Code, w.Body.String())
+		}
+		if mock.CompletionRequest.Prompt != "compare [img-0] and [img-1]" {
+			t.Fatalf("completion prompt = %q, want image marker prompt", mock.CompletionRequest.Prompt)
+		}
+		if len(mock.CompletionRequest.Media) != 2 {
+			t.Fatalf("completion media = %#v, want two images", mock.CompletionRequest.Media)
+		}
+		if mock.CompletionRequest.Media[0].ID != 0 || !bytes.Equal(mock.CompletionRequest.Media[0].Data, image1) {
+			t.Fatalf("first image media = %#v", mock.CompletionRequest.Media[0])
+		}
+		if mock.CompletionRequest.Media[1].ID != 1 || !bytes.Equal(mock.CompletionRequest.Media[1].Data, image2) {
+			t.Fatalf("second image media = %#v", mock.CompletionRequest.Media[1])
+		}
+	})
+}
+
 func TestGenerateChatRemote(t *testing.T) {
 	gin.SetMode(gin.TestMode)