diff --git a/internal/entity/models/openai.go b/internal/entity/models/openai.go
index 46b52df406..9f17895be0 100644
--- a/internal/entity/models/openai.go
+++ b/internal/entity/models/openai.go
@@ -124,10 +124,14 @@ func (o *OpenAIModel) ChatWithMessages(modelName string, messages []Message, api
 		}
 	}
 
-	// Qwen3 family: disable thinking by default (matches Python's
-	// _apply_model_family_policies in rag/llm/chat_model.py:119-121).
-	if strings.Contains(strings.ToLower(modelName), "qwen3") && (chatModelConfig == nil || chatModelConfig.Thinking == nil) {
-		reqBody["enable_thinking"] = false
+	// Qwen3 family: disable thinking unless explicitly enabled (matches
+	// Python's _apply_model_family_policies in rag/llm/chat_model.py:119-121).
+	if strings.Contains(strings.ToLower(modelName), "qwen3") {
+		if chatModelConfig != nil && chatModelConfig.Thinking != nil {
+			reqBody["enable_thinking"] = *chatModelConfig.Thinking
+		} else {
+			reqBody["enable_thinking"] = false
+		}
 	}
 
 	jsonData, err := json.Marshal(reqBody)
@@ -302,9 +306,13 @@ func (o *OpenAIModel) ChatStreamlyWithSender(modelName string, messages []Messag
 		}
 	}
 
-	// Qwen3 family: disable thinking by default.
-	if strings.Contains(strings.ToLower(modelName), "qwen3") && (chatModelConfig == nil || chatModelConfig.Thinking == nil) {
-		reqBody["enable_thinking"] = false
+	// Qwen3 family: disable thinking unless explicitly enabled.
+	if strings.Contains(strings.ToLower(modelName), "qwen3") {
+		if chatModelConfig != nil && chatModelConfig.Thinking != nil {
+			reqBody["enable_thinking"] = *chatModelConfig.Thinking
+		} else {
+			reqBody["enable_thinking"] = false
+		}
 	}
 
 	jsonData, err := json.Marshal(reqBody)
diff --git a/internal/service/ask_service.go b/internal/service/ask_service.go
index c5c1fbe008..40ef84a453 100644
--- a/internal/service/ask_service.go
+++ b/internal/service/ask_service.go
@@ -186,7 +186,12 @@ func (s *AskService) run(ctx context.Context, llm StreamingLLM, userID, question
 		{Role: "system", Content: sysPrompt},
 		{Role: "user", Content: question},
 	}
-	genConf := &modelModule.ChatConfig{Temperature: ptrFloat64(0.1)}
+	// Thinking is disabled: summarization does not need reasoning. With
+	// thinking enabled, the reasoning (which drafts the summary) streams
+	// first, then collapses into a hidden think block, and the provider may
+	// emit only a fragment as the visible answer once the reasoning has
+	// consumed the output budget.
+	genConf := &modelModule.ChatConfig{Temperature: ptrFloat64(0.1), Thinking: ptrBool(false)}
 
 	ch, err := llm.ChatStream(ctx, messages, genConf)
 	if err != nil {
@@ -279,3 +284,4 @@ func toFloat64Slice(v interface{}) []float64 {
 
 func ptrInt(v int) *int             { return &v }
 func ptrFloat64(v float64) *float64 { return &v }
+func ptrBool(v bool) *bool          { return &v }