[Bug]: Qwen3.5 Can not close thinking by "enable_thinking": false

[Charmnut]

Your current environment

Details

The vllm/vllm-openai:qwen3_5 image is deployed

🐛 Describe the bug

Adding "chat_template_kwargs": {"enable_thinking": false} to the HTTP body does not disable the thinking process.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[chaunceyjiang]

The main branch already supports this — please wait for the update.

[chegewara]

I am having issue to run qwen3.5 on release 0.16.0, but i can run qwen3.5-27B-fp8 and qwen3.5-30b-fp8 on 2x3090 installed with uv nightly.
0.16.1rc1.dev85+g90805ff46

This vllm version works great with all params, including reasoning control from payload.

PS i mean i am using vllm serve, not python vllm or docker.

[Charmnut]

The main branch already supports this — please wait for the update.

I have tried today's nightly image, it doesn't work

[ubbiiy]

The main branch already supports this — please wait for the update.

I have tried today's nightly image, it doesn't work

I downloaded the nightly build yesterday and it works fine. So I suppose it should work in today's one too.

[Charmnut]

The main branch already supports this — please wait for the update.

I have tried today's nightly image, it doesn't work

I downloaded the nightly build yesterday and it works fine. So I suppose it should work in today's one too.

Could you tell me your method of closing qwen3.5 thinking?

[315930399]

The main branch already supports this — please wait for the update.

I have tried today's nightly image, it doesn't work

I downloaded the nightly build yesterday and it works fine. So I suppose it should work in today's one too.

Could you tell me your method of closing qwen3.5 thinking?

我试了可以的啊

[Charmnut]

Image:
docker pull vllm/vllm-openai:nightly-afd089f231d714e7fd06b51e3bc7df7fe004c7f9

Deploy command：
vllm serve /mnt/models/Qwen3.5-397B-A17B-FP8 --trust-remote-code --tensor-parallel-size 8 --served-model-name qwen3_fp8 --port 1025 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

Post body

{
    "model": "qwen3_fp8",
     "messages": [{
        "role": "user",
        "content": "你好你是谁"
    }],
    "stream": false,
    "details": false,
    "max_tokens": 8192,
    "chat_template_kwargs": {"enable_thinking": false}
}

Anyone know where I went wrong?

[chegewara]

request from docker, thus odd url (sorry):

curl -X POST http: //host.docker.internal:8000/v1/chat/completions   -H "Content-Type: application/json"   -d '{
"model": "qwen3.5-35B",
"messages": [
    {
        "role": "user",
        "content": "hi"
    }
],
"max_tokens": 32768,
"temperature": 0.7,
"top_p": 0.8,
"presence_penalty": 1.5,
"seed": 642832912,
"chat_template_kwargs": {
    "enable_thinking": false
},
"stream": false, # works both, false/true
"tools": []
}'

response (no reasoning/thinking):

{
"id": "chatcmpl-87be511cf20c5134",
"object": "chat.completion",
"created": 1772425252,
"model": "qwen3.5-35B",
"choices": [
    {
        "index": 0,
        "message": {
            "role": "assistant",
            "content": "Hello! How can I help you today?",
            "refusal": null,
            "annotations": null,
            "audio": null,
            "function_call": null,
            "tool_calls": [],
            "reasoning": null
        },
        "logprobs": null,
        "finish_reason": "stop",
        "stop_reason": null,
        "token_ids": null
    }
],
"service_tier": null,
"system_fingerprint": null,
"usage": {
    "prompt_tokens": 13,
    "total_tokens": 23,
    "completion_tokens": 10,
    "prompt_tokens_details": null
},
"prompt_logprobs": null,
"prompt_token_ids": null,
"kv_transfer_params": null
}

vllm version 0.16.1rc1.dev85+g90805ff46

vllm serve Qwen/Qwen3.5-35B-A3B-FP8 --trust-remote-code --host 0.0.0.0 --port 8000  --enable-force-include-usage --tensor-parallel-size 2 --max-model-len 200k --gpu-memory-utilization 0.93 --kv-cache-dtype fp8 --enable-auto-tool-choice --tool-call-parser qwen3_coder --max-num-batched-tokens 4096 --enforce-eager --served-model-name qwen3     --reasoning-parser qwen3 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":4}'

PS MTP nicely speed up inference, i love it

[liangpn]

request from docker, thus odd url (sorry):

curl -X POST http: //host.docker.internal:8000/v1/chat/completions   -H "Content-Type: application/json"   -d '{
"model": "qwen3.5-35B",
"messages": [
    {
        "role": "user",
        "content": "hi"
    }
],
"max_tokens": 32768,
"temperature": 0.7,
"top_p": 0.8,
"presence_penalty": 1.5,
"seed": 642832912,
"chat_template_kwargs": {
    "enable_thinking": false
},
"stream": false, # works both, false/true
"tools": []
}'

response (no reasoning/thinking):

{
"id": "chatcmpl-87be511cf20c5134",
"object": "chat.completion",
"created": 1772425252,
"model": "qwen3.5-35B",
"choices": [
    {
        "index": 0,
        "message": {
            "role": "assistant",
            "content": "Hello! How can I help you today?",
            "refusal": null,
            "annotations": null,
            "audio": null,
            "function_call": null,
            "tool_calls": [],
            "reasoning": null
        },
        "logprobs": null,
        "finish_reason": "stop",
        "stop_reason": null,
        "token_ids": null
    }
],
"service_tier": null,
"system_fingerprint": null,
"usage": {
    "prompt_tokens": 13,
    "total_tokens": 23,
    "completion_tokens": 10,
    "prompt_tokens_details": null
},
"prompt_logprobs": null,
"prompt_token_ids": null,
"kv_transfer_params": null
}

vllm version 0.16.1rc1.dev85+g90805ff46

vllm serve Qwen/Qwen3.5-35B-A3B-FP8 --trust-remote-code --host 0.0.0.0 --port 8000  --enable-force-include-usage --tensor-parallel-size 2 --max-model-len 200k --gpu-memory-utilization 0.93 --kv-cache-dtype fp8 --enable-auto-tool-choice --tool-call-parser qwen3_coder --max-num-batched-tokens 4096 --enforce-eager --served-model-name qwen3     --reasoning-parser qwen3 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":4}'

PS MTP nicely speed up inference, i love it

docker pull vllm/vllm-openai:nightly-afd089f231d714e7fd06b51e3bc7df7fe004c7f9？

[chaunceyjiang]

PS MTP nicely speed up inference, i love it

see #35615

[Charmnut]

I didn’t get this issue solved until I switched to SGLang, so I don’t think it was a problem with how I deployed or operated it. Sometimes, a smaller open-source project isn’t necessarily a bad thing.

[blockoracle]

Fixes 1 and 19 in my Qwen 3.5 chat template address this:

• Fix 1: Safe defaults for enable_thinking (defaults to true, properly handles undefined)
• Fix 19: Auto-disable thinking when tools are provided (configurable via auto_disable_thinking_with_tools)

Template: https://huggingface.co/barubary/qwen3.5-barubary-attuned-chat-template

[ValeKnappich]

@blockoracle why would you auto-disable thinking when tools are provided?