honor explicit thinking_enable=False on Ollama and Gemini

Author: qbc2016

scripts/model_examples/openai_chat_call.py

@@ -34,6 +34,7 @@ async def example_simple_call() -> None:
         stream=True,
         context_size=200_000,
         parameters=OpenAIChatModel.Parameters(
+            thinking_enable=True,
             reasoning_effort="low",
         ),
     )
@@ -80,6 +81,7 @@ async def example_tool_call() -> None:
         stream=True,
         context_size=200_000,
         parameters=OpenAIChatModel.Parameters(
+            thinking_enable=True,
             reasoning_effort="low",
         ),
     )
@@ -155,6 +157,7 @@ async def example_structured_output() -> None:
         stream=True,
         context_size=200_000,
         parameters=OpenAIChatModel.Parameters(
+            thinking_enable=True,
             reasoning_effort="low",
         ),
     )

scripts/model_examples/openai_response_call.py

@@ -38,7 +38,10 @@ async def example_simple_call() -> None:
         model="o1",
         stream=True,
         context_size=200_000,
-        parameters=OpenAIResponseModel.Parameters(reasoning_effort="low"),
+        parameters=OpenAIResponseModel.Parameters(
+            thinking_enable=True,
+            reasoning_effort="low",
+        ),
     )
 
     msgs = [
@@ -82,7 +85,10 @@ async def example_tool_call() -> None:
         model="o1",
         stream=True,
         context_size=200_000,
-        parameters=OpenAIResponseModel.Parameters(reasoning_effort="low"),
+        parameters=OpenAIResponseModel.Parameters(
+            thinking_enable=True,
+            reasoning_effort="low",
+        ),
     )
 
     msgs = [
@@ -158,7 +164,10 @@ async def example_structured_output() -> None:
         model="o1",
         stream=True,
         context_size=200_000,
-        parameters=OpenAIResponseModel.Parameters(reasoning_effort="low"),
+        parameters=OpenAIResponseModel.Parameters(
+            thinking_enable=True,
+            reasoning_effort="low",
+        ),
     )
 
     msgs = [

scripts/model_examples/xai_call.py

@@ -38,7 +38,10 @@ async def example_simple_call() -> None:
         model="grok-3-mini",
         stream=True,
         context_size=131_072,
-        parameters=XAIChatModel.Parameters(reasoning_effort="low"),
+        parameters=XAIChatModel.Parameters(
+            thinking_enable=True,
+            reasoning_effort="low",
+        ),
     )
 
     msgs = [
@@ -157,7 +160,10 @@ async def example_structured_output() -> None:
         model="grok-3-mini",
         stream=True,
         context_size=131_072,
-        parameters=XAIChatModel.Parameters(reasoning_effort="low"),
+        parameters=XAIChatModel.Parameters(
+            thinking_enable=True,
+            reasoning_effort="low",
+        ),
     )
 
     msgs = [

src/agentscope/model/_gemini/_model.py

@@ -89,10 +89,13 @@ class Parameters(BaseModel):
             gt=0,
         )
 
-        thinking_enable: bool = Field(
-            default=False,
+        thinking_enable: bool | None = Field(
+            default=None,
             title="Thinking",
-            description="Whether to enable thinking output.",
+            description="Whether to enable thinking output."
+            " ``None`` keeps the Gemini model's own default"
+            " (2.5 Pro/Flash think by default);"
+            " ``True``/``False`` explicitly turns thinking on/off.",
         )
 
         thinking_budget: int | None = Field(
@@ -234,11 +237,17 @@ async def _call_api(
         if self.parameters.top_p is not None:
             config["top_p"] = self.parameters.top_p
 
-        if self.parameters.thinking_enable:
-            config["thinking_config"] = {
-                "include_thoughts": True,
-                "thinking_budget": self.parameters.thinking_budget or 1024,
-            }
+        if self.parameters.thinking_enable is not None:
+            if self.parameters.thinking_enable:
+                config["thinking_config"] = {
+                    "include_thoughts": True,
+                    "thinking_budget": self.parameters.thinking_budget or 1024,
+                }
+            else:
+                config["thinking_config"] = {
+                    "include_thoughts": False,
+                    "thinking_budget": 0,
+                }
 
         fmt_tools, fmt_tool_choice = self._format_tools(tools, tool_choice)
 

src/agentscope/model/_ollama/_model.py

@@ -35,11 +35,13 @@ class Parameters(BaseModel):
             gt=0,
         )
 
-        thinking_enable: bool = Field(
-            default=False,
+        thinking_enable: bool | None = Field(
+            default=None,
             title="Thinking",
             description="Whether to enable thinking"
-            " (for models like qwen3, deepseek-r1).",
+            " (for models like qwen3, deepseek-r1)."
+            " ``None`` keeps the Ollama model's own default;"
+            " ``True``/``False`` explicitly turns thinking on/off.",
         )
 
         temperature: float | None = Field(
@@ -175,8 +177,8 @@ async def _call_api(
         if options:
             kwargs["options"] = options
 
-        if self.parameters.thinking_enable:
-            kwargs["think"] = True
+        if self.parameters.thinking_enable is not None:
+            kwargs["think"] = self.parameters.thinking_enable
 
         kwargs.update(generate_kwargs)