qwen3.6-27b - cannot disable thinking via OpenAI API endpoint

[jackwh]

Hey there! Firstly oMLX is awesome, thank you for making MLX models so easy to run! 😎

Describe the bug

I'm calling mlx-community/Qwen3.6-27B-mxfp8 via the OpenAI API endpoint. No matter what I try, I can't get API request parameters to disable thinking.

Thinking can be disabled in the oMLX model settings by changing the preset, but I'd like to be able to do it programmatically at runtime.

The Qwen 3.6 model card says to set "chat_template_kwargs": {"enable_thinking": False}, in the request. I've tried this - as well as just a root-level enable_thinking parameter, not nested under chat_template_kwags but alongside parameters like temperature etc. - but it makes no difference.

oMLX logs show the parameters are being received, it just seems it's not affecting the model output.

I assume this is possible, but couldn't find docs explaining what I'm meant to be doing (if it's something different).

Expected behavior

Model thinking should be able to be disabled at runtime by API call, not just in the UI model settings.

Desktop (please complete the following information):

• macOS Version: Tahoe 26.4.1
• oMLX Version 0.3.9

[jundot]

Thanks for the kind words and for the report!

I just reproduced this against Qwen3.6-27B on 0.3.9, and chat_template_kwargs: {"enable_thinking": false} does work as expected on the OpenAI endpoint. Here's the side-by-side:

thinking ON (default for Qwen 3.6):

curl -s http://localhost:8000/v1/chat/completions \
  -H "Authorization: Bearer test1234" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3.6-27B-oQ5-fp16",
    "messages": [{"role":"user","content":"Name the capital of France in one word."}],
    "max_tokens": 300,
    "chat_template_kwargs": {"enable_thinking": true}
  }'
# prompt_tokens=19, completion_tokens=158, content="Paris", reasoning_content=536 chars

thinking OFF:

curl -s http://localhost:8000/v1/chat/completions \
  -H "Authorization: Bearer test1234" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3.6-27B-oQ5-fp16",
    "messages": [{"role":"user","content":"Name the capital of France in one word."}],
    "max_tokens": 300,
    "chat_template_kwargs": {"enable_thinking": false}
  }'
# prompt_tokens=21, completion_tokens=1, content="Paris", reasoning_content=""

Note prompt_tokens going 19 → 21: that's the Qwen template inserting the empty <think></think> block when enable_thinking is false, which is the direct evidence the flag reached the tokenizer.

A couple of things to check on your side:

1. enable_thinking must be a JSON boolean (false, lowercase, no quotes) nested under chat_template_kwargs. A top-level enable_thinking is silently dropped because OpenAI's spec doesn't define that field. I verified this too: top-level false produced the same thinking-ON output as default.
2. In oMLX's per-model settings, enable_thinking and its "force" lock are separate toggles. Setting enable_thinking=true does not auto-force the value. The force lock is only set if you explicitly tick the "force" box next to enable_thinking. If you have the force lock on, per-request chat_template_kwargs for that key is intentionally ignored so the model stays in the configured mode. Unchecking it lets API requests override.

If neither of those is the cause, please share the exact request body (or the SDK call) you're using and I'll dig further into the mxfp8 path specifically.