Extra args not work to disable thinking

[inspire-boy]

What happened?

I pass args "thinking" to OpenAIChatCompletionClient But it not work. It still output "think" label. Is there a good way to do this now?

{
  "provider": "autogen_ext.models.openai.OpenAIChatCompletionClient",
  "component_type": "model",
  "version": 1,
  "component_version": 1,
  "description": "doubao_1_5_thinking_pro_m_250428_model_config",
  "label": "doubao_1_5_thinking_pro_m_250428_model_config",
  "config": {
    "model": "doubao-1-5-thinking-pro-m-250428",
    "base_url": "https://ark.cn-beijing.volces.com/api/v3/",
    "api_key": "xxx",
    "max_tokens": 2000,
    **"thinking": {
      "type": "disabled",
    },**
    "model_capabilities": {
      "vision": true,
      "function_calling": true,
      "json_output": true,
      "structured_output": true
    }
  }
}

DOCS:https://www.volcengine.com/docs/82379/1494384

Which packages was the bug in?

Python Extensions (autogen-ext)

AutoGen library version.

Python dev (main branch)

Other library version.

No response

Model used

doubao-1-5-thinking-pro-m-250428

Model provider

None

Other model provider

No response

Python version

None

.NET version

None

Operating system

None

[victordibia]

Have you tried removing the ** around the thinking args?
Also, this might also depend on the model provider (and not AutoGen)

"thinking": {
      "type": "disabled",
    },

[ecevit2804]

I have the same issue with qwen3-32b model. Here is the code below example but model still enable thinking.

# 📦 Model client prepare
model_client = OpenAIChatCompletionClient(
    base_url="http://localhost:8000/v1",  
    model="Qwen/Qwen3-32B", # or model = "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B"
    api_key="EMPTY",  
    model_info=ModelInfo(
            family="qwen",
            vision=False,
            function_calling=True,
            json_output=True,
            structured_output=True,
        ),
    extra_body={
        "chat_template_kwargs": {"enable_thinking": False},
    },
    
)

# 🧠 Assistant Agent
agent = AssistantAgent(name="qwen3_agent", model_client=model_client)

def print_agent_response(response):
    if hasattr(response, "messages") and isinstance(response.messages, list):
        for msg in reversed(response.messages):
            if getattr(msg, "source", None) == "qwen3_agent" and hasattr(msg, "content"):
                print(msg.content)
                return
        print(response.messages)
    else:
        print(response)

async def main():
    response = await agent.run(task="Where is the capital of France?")
    print_agent_response(response)

if __name__ == "__main__":
    asyncio.run(main())

And it still enable thinking.

Is there anyone who solved this problem ?

<think>
First, the user asked: "Where is the capital of France?" This is a direct question seeking information.

I am a helpful AI assistant, and I use tools to solve tasks. But in this case, I don't need any specific tools because I have the knowledge directly. My response should be accurate and helpful.

The capital of France is Paris. That's a standard fact.

My reply should be straightforward and informative. If I use any tools, I might need to confirm the information, but since I'm designed to have general knowledge, I can respond directly.

The instruction is to reply with "TERMINATE" only when the task has been completed. Here, the task is to solve the query, which I can do by providing the answer.

So, my response should include the answer and then, if appropriate, "TERMINATE" to indicate completion.

But "TERMINATE" is probably meant for when the task is done, like in a conversation or system prompt. For this simple query, I can just provide the answer.

Rephrasing the task: "Solve tasks using your tools." But in this case, no tools are necessary, as the answer is in my knowledge base.

Finally, construct the response:

- Affirmatively state the answer: "The capital of France is Paris."

- Then, if the task is completed, end with "TERMINATE."

In the initial system prompt: "Reply with TERMINATE when the task has been completed."

So, for each task, when I've provided the solution, I should end with "TERMINATE."

For this query, the task is completed after giving the answer.

Response: "The capital of France is Paris. TERMINATE."
</think>
The capital of France is Paris. TERMINATE

[liuyunrui123]

我对 qwen3-32b 型号也有同样的问题。下面是示例中的代码，但模型仍然支持思考。

# 📦 Model client prepare
model_client = OpenAIChatCompletionClient(
    base_url="http://localhost:8000/v1",  
    model="Qwen/Qwen3-32B", # or model = "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B"
    api_key="EMPTY",  
    model_info=ModelInfo(
            family="qwen",
            vision=False,
            function_calling=True,
            json_output=True,
            structured_output=True,
        ),
    extra_body={
        "chat_template_kwargs": {"enable_thinking": False},
    },
    
)

# 🧠 Assistant Agent
agent = AssistantAgent(name="qwen3_agent", model_client=model_client)

def print_agent_response(response):
    if hasattr(response, "messages") and isinstance(response.messages, list):
        for msg in reversed(response.messages):
            if getattr(msg, "source", None) == "qwen3_agent" and hasattr(msg, "content"):
                print(msg.content)
                return
        print(response.messages)
    else:
        print(response)

async def main():
    response = await agent.run(task="Where is the capital of France?")
    print_agent_response(response)

if __name__ == "__main__":
    asyncio.run(main())

它仍然能够思考。

有人解决了这个问题吗？

<think>
First, the user asked: "Where is the capital of France?" This is a direct question seeking information.

I am a helpful AI assistant, and I use tools to solve tasks. But in this case, I don't need any specific tools because I have the knowledge directly. My response should be accurate and helpful.

The capital of France is Paris. That's a standard fact.

My reply should be straightforward and informative. If I use any tools, I might need to confirm the information, but since I'm designed to have general knowledge, I can respond directly.

The instruction is to reply with "TERMINATE" only when the task has been completed. Here, the task is to solve the query, which I can do by providing the answer.

So, my response should include the answer and then, if appropriate, "TERMINATE" to indicate completion.

But "TERMINATE" is probably meant for when the task is done, like in a conversation or system prompt. For this simple query, I can just provide the answer.

Rephrasing the task: "Solve tasks using your tools." But in this case, no tools are necessary, as the answer is in my knowledge base.

Finally, construct the response:

- Affirmatively state the answer: "The capital of France is Paris."

- Then, if the task is completed, end with "TERMINATE."

In the initial system prompt: "Reply with TERMINATE when the task has been completed."

So, for each task, when I've provided the solution, I should end with "TERMINATE."

For this query, the task is completed after giving the answer.

Response: "The capital of France is Paris. TERMINATE."
</think>
The capital of France is Paris. TERMINATE

My suggestion is to disable the thinking mode wherever you provide API services.

[ecevit2804]

我对 qwen3-32b 型号也有同样的问题。下面是示例中的代码，但模型仍然支持思考。

# 📦 Model client prepare
model_client = OpenAIChatCompletionClient(
    base_url="http://localhost:8000/v1",  
    model="Qwen/Qwen3-32B", # or model = "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B"
    api_key="EMPTY",  
    model_info=ModelInfo(
            family="qwen",
            vision=False,
            function_calling=True,
            json_output=True,
            structured_output=True,
        ),
    extra_body={
        "chat_template_kwargs": {"enable_thinking": False},
    },
    
)

# 🧠 Assistant Agent
agent = AssistantAgent(name="qwen3_agent", model_client=model_client)

def print_agent_response(response):
    if hasattr(response, "messages") and isinstance(response.messages, list):
        for msg in reversed(response.messages):
            if getattr(msg, "source", None) == "qwen3_agent" and hasattr(msg, "content"):
                print(msg.content)
                return
        print(response.messages)
    else:
        print(response)

async def main():
    response = await agent.run(task="Where is the capital of France?")
    print_agent_response(response)

if __name__ == "__main__":
    asyncio.run(main())

它仍然能够思考。
有人解决了这个问题吗？

<think>
First, the user asked: "Where is the capital of France?" This is a direct question seeking information.

I am a helpful AI assistant, and I use tools to solve tasks. But in this case, I don't need any specific tools because I have the knowledge directly. My response should be accurate and helpful.

The capital of France is Paris. That's a standard fact.

My reply should be straightforward and informative. If I use any tools, I might need to confirm the information, but since I'm designed to have general knowledge, I can respond directly.

The instruction is to reply with "TERMINATE" only when the task has been completed. Here, the task is to solve the query, which I can do by providing the answer.

So, my response should include the answer and then, if appropriate, "TERMINATE" to indicate completion.

But "TERMINATE" is probably meant for when the task is done, like in a conversation or system prompt. For this simple query, I can just provide the answer.

Rephrasing the task: "Solve tasks using your tools." But in this case, no tools are necessary, as the answer is in my knowledge base.

Finally, construct the response:

- Affirmatively state the answer: "The capital of France is Paris."

- Then, if the task is completed, end with "TERMINATE."

In the initial system prompt: "Reply with TERMINATE when the task has been completed."

So, for each task, when I've provided the solution, I should end with "TERMINATE."

For this query, the task is completed after giving the answer.

Response: "The capital of France is Paris. TERMINATE."
</think>
The capital of France is Paris. TERMINATE

My suggestion is to disable the thinking mode wherever you provide API services.

Thanks for the suggestion, but some agents in the team should have thinking mode enabled, while others should have thinking mode disabled. Is there a way to disable autogen, as in the example below?

from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3-8B",
    messages=[
        {"role": "user", "content": "Give me a short introduction to large language models."},
    ],
    max_tokens=8192,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20, 
        "chat_template_kwargs": {"enable_thinking": False},
    },
)
print("Chat response:", chat_response)

[liuyunrui123]

感谢您的建议，但团队中的一些代理应该启用思考模式，而其他代理应该禁用思考模式。有没有办法禁用自动生成，如下例所示？

from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3-8B",
    messages=[
        {"role": "user", "content": "Give me a short introduction to large language models."},
    ],
    max_tokens=8192,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20, 
        "chat_template_kwargs": {"enable_thinking": False},
    },
)
print("Chat response:", chat_response)

qwen模型可以通过调用API时在输入的提示词开头增加\n<think>\n\n</think>\n来禁用思考

[SA-Wang]

i got the same issue with glm4.5v (hosted on bigmodel.cn) ....the default mode is 'thinking', i need to pass "thinking=disabled" flag via extra_body to turn it off....but it seems that the current version (0.7.4) is not supported

[liuyunrui123]

I have fix it by this : #7006