server: apply format constraint for all thinking parsers when think=false (#15901)

server/routes.go

@@ -2766,9 +2766,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			// current approach uses the transition from parsed thinking content to
 			// parsed non-thinking content as the signal to turn constraining on
 
-			// TODO(parthsareen): temporary fix for https://github.com/ollama/ollama/issues/15260.
-			// To revisit for other models and have a consistent pattern across models through parsers.
-			forceImmediate := m.Config.Parser == "gemma4" && req.Think != nil && !req.Think.Bool()
+			forceImmediate := builtinParser != nil && builtinParser.HasThinkingSupport() && req.Think != nil && !req.Think.Bool()
 			if req.Format != nil && structuredOutputsState == structuredOutputsState_None && !forceImmediate && ((builtinParser != nil || thinkingState != nil) && slices.Contains(m.Capabilities(), model.CapabilityThinking)) {
 				currentFormat = nil
 			}

server/routes_generate_test.go

@@ -2811,15 +2811,18 @@ func TestChatWithPromptEndingInThinkTag(t *testing.T) {
 }
 
 // TestChatFormatWithThinkFalse verifies that when a model uses a builtin
-// parser that supports thinking (e.g. gemma4) and the request explicitly
-// disables thinking (think=false), the format constraint is passed to the
-// first and only completion call. Previously, format was deferred for all
-// thinking-capable parsers and only re-applied after an end-of-thinking
-// transition — a transition that never fires when thinking is off. See
-// https://github.com/ollama/ollama/issues/15260.
+// parser that supports thinking and the request explicitly disables thinking
+// (think=false), the format constraint is passed to the first and only
+// completion call. Previously, format was deferred for all thinking-capable
+// parsers and only re-applied after an end-of-thinking transition -- a
+// transition that never fires when thinking is off. See
+// https://github.com/ollama/ollama/issues/15260 and
+// https://github.com/ollama/ollama/issues/14645.
 func TestChatFormatWithThinkFalse(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 
+	for _, parserName := range []string{"gemma4", "qwen3.5", "qwen3-thinking"} {
+		t.Run(parserName, func(t *testing.T) {
 			mock := &mockRunner{
 				CompletionResponse: llm.CompletionResponse{
 					Done:               true,
@@ -2876,13 +2879,11 @@ func TestChatFormatWithThinkFalse(t *testing.T) {
 				{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
 			})
 
-	// Use the gemma4 builtin parser — it reports HasThinkingSupport=true, which
-	// adds CapabilityThinking to the model and previously triggered deferral of
-	// the format even when the user passed think=false.
+			modelName := "test-" + parserName + "-parser"
 			w := createRequest(t, s.CreateHandler, api.CreateRequest{
-		Model:    "test-gemma4-parser",
+				Model:    modelName,
 				Files:    map[string]string{"file.gguf": digest},
-		Parser:   "gemma4",
+				Parser:   parserName,
 				Template: `{{- range .Messages }}{{ .Role }}: {{ .Content }}{{ end }}`,
 				Stream:   &stream,
 			})
@@ -2916,7 +2917,7 @@ func TestChatFormatWithThinkFalse(t *testing.T) {
 			streamRequest := false
 			think := false
 			w = createRequest(t, s.ChatHandler, api.ChatRequest{
-		Model:    "test-gemma4-parser",
+				Model:    modelName,
 				Messages: []api.Message{{Role: "user", Content: "Respond in JSON."}},
 				Think:    &api.ThinkValue{Value: think},
 				Stream:   &streamRequest,
@@ -2934,6 +2935,8 @@ func TestChatFormatWithThinkFalse(t *testing.T) {
 			if !bytes.Equal([]byte(format), []byte(requests[0].Format)) {
 				t.Errorf("expected first completion format to match the request format, got %q", string(requests[0].Format))
 			}
+		})
+	}
 }
 
 func TestGenerateUnload(t *testing.T) {