Unable to close QWen3's thinks mode in LMStudio.

[yuluod]

In the API, I can set enable_thinking=False to turn it off, but it seems I didn't see this option in LMStudo.

[wqerrewetw]

/no_think

[yieme]

I think wqerrewetw is right: /no_think to stop thinking and /think to start thinking (default). I add to the top of my system prompt. note, if multiples of these are used, the last one is what seems to stick, ex: /no_think later followed by /think will be back into think mode. the 0.6B model version seems to not observe /no_think and so it always thinks.

[wqw547243068]

/no_think did not work

[hungrymonkey]

I cannot figure it out.

curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-0.6b-mlx",
    "messages": [
      { "role": "system", "content": "Always answer in rhymes. Today is Thursday" },
      { "role": "user", "content": "What day is it today?" }
    ],
    "temperature": 0.7,
    "max_tokens": -1,
    "stream": false,
  "chat_template_kwargs": {"enable_thinking": false}
}'
{
  "id": "chatcmpl-foeuhndpoap20kg595i0y9",
  "object": "chat.completion",
  "created": 1756479438,
  "model": "qwen3-0.6b-mlx",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "<think>\nOkay, the user asked, \"What day is it today?\" and I need to answer in rhymes. Let me think.\n\nThe user mentioned today is Thursday, so the response should be a rhyme with \"Thursday\" and maybe something about that day. Let me brainstorm some rhymes. Maybe start with \"Thursday, Thursday\" and then include a rhyme about being a day or something related to the day. Words like \"lantern\" or \"clock\" could work. Let me check syllable counts to make sure it's easy and rhymes. \"Thursday, Thursday, night is coming,\" that's two lines with a rhyme. The second line rhymes with \"night\" and \"clock.\" That should fit. I'll go with that.\n</think>\n\nThursday, Thursday, night is coming!  \nWith a clock and a lantern, we’re dancing.",
        "tool_calls": []
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 28,
    "completion_tokens": 176,
    "total_tokens": 204
  },
  "stats": {},
  "system_fingerprint": "qwen3-0.6b-mlx"
}

curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-0.6b-mlx",
    "messages": [
      { "role": "system", "content": "Always answer in rhymes. Today is Thursday" },
      { "role": "user", "content": "What day is it today?" }
    ],
    "temperature": 0.7,
    "max_tokens": -1,
    "stream": false,
  "enable_thinking": "False"
}'
{
  "id": "chatcmpl-ziaur8xlfl8phc73ch094",
  "object": "chat.completion",
  "created": 1756479584,
  "model": "qwen3-0.6b-mlx",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "<think>\nOkay, the user asked \"Today is Thursday.\" I need to respond in a way that rhymes. Let me think... The answer is Thursday, so maybe start with \"Thursday!\" and then add some rhyming lines. Let me try:\n\n\"Thursday, Thursday! I'm here to tell you today is the day we're all going to have fun.\" That works because \"Thursday\" and \"today\" rhyme, and the rest of the lines follow. Let me check again for rhythm. Each line should be a couplet, so \"Thursday! I'm here to tell you today is the day we're all going to have fun.\" That's correct. I think that's it.\n</think>\n\nThursday!  \nI'm here to tell you today is the day we're all going to have fun.",
        "tool_calls": []
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 28,
    "completion_tokens": 166,
    "total_tokens": 194
  },
  "stats": {},
  "system_fingerprint": "qwen3-0.6b-mlx"
}

[saaay71]

In LM studio, go to the Developer tap -> Inference -> Custom Fields -> Enable Thinking
This way you disable thinking completely.

[sigmareaver]

In LM studio, go to the Developer tap -> Inference -> Custom Fields -> Enable Thinking This way you disable thinking completely.

I'm not seeing a custom fields section in Developer -> Inference on the new UI (LMStudio 0.4.2). Was it moved or removed?

Only thing I've found in the docs is this:

curl http://localhost:1234/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-oss-20b",
    "input": "Provide a prime number less than 50",
    "reasoning": { "effort": "low" }
  }'

But... this isn't useful if you're using the lmstudio Python package.

[Iory1998]

I'm not seeing a custom fields section in Developer -> Inference on the new UI (LMStudio 0.4.2). Was it moved or removed?

The issue is that only remove the thinking parser not the turning off the thinking mode.

[heby1]

I found this was hidden under Custom Fields.

[LiJT]

I found this was hidden under Custom Fields.

I dont have this custom field section for my model Qwen3.5-27B-Uncensored-HauhauCS-Aggressive

[heby1]

I dont have this custom field section for my model Qwen3.5-27B-Uncensored-HauhauCS-Aggressive

Actually, even setting the Custom Fields Enable Thinking = off on 9B model did not disable thinking for me.

This however worked and disabled thinking on 27B model, probably will also work on other Qwen3.5 models:

Go to My Models > Inference > Prompt Template and change it to ChatML. I don't know if there are better versions available. Vision functionality also still works so it seems good enough.

[LiJT]

I dont have this custom field section for my model Qwen3.5-27B-Uncensored-HauhauCS-Aggressive

Actually, even setting the Custom Fields Enable Thinking = off on 9B model did not disable thinking for me.

This however worked and disabled thinking on 27B model, probably will also work on other Qwen3.5 models:

Go to My Models > Inference > Prompt Template and change it to ChatML. I don't know if there are better versions available. Vision functionality also still works so it seems good enough.

Wow thanks man!
Also I found a way to trick LM studio and get the Thinking button.
First
Navigate to

C:\Users\[Your Username]\.lmstudio\hub\models\qwen\
(Use qwen3.5 for example)

Create a folder after your model name, mine is Qwen3.5-27B-Uncensored-HauhauCS-Aggressive so

C:\Users\charl\.lmstudio\hub\models\qwen\qwen3.5-27b-uncensored-hauhaucs-aggressive

then drop these two files inside that folder,
manifest.json
model.yaml

These two files are for Qwen3.5-27B-Uncensored-HauhauCS-Aggressive only! You need to change several lines for your own unoffcial model to work!

Then you will notice the thinking function is unlocked for your model

If you have other model, change these files accordingly, then LM studio will think you downloaded the original stuff picked model and enables the think button for you

I really wish LM studio comes up with an automatic way to generate these files for ease of use

[heby1]

That is a very neat and better way for sure, making it easy to switch from thinking to non-thinking!

I noticed these files are present in another original Qwen model:
C:\Users\xxxx\.cache\lm-studio\hub\models\qwen\qwen3.5-35b-a3b

Both manifest.json and model.yaml are there. But those are not present with other official Qwen model releases, so I guess we're still tinkering ...

[DrClairvoyant]

How do I do this with the Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive model? This method works for the 9B model. I tried altered some lines as respected but it didn't work.

FYI I did install the official Qwen 9B model with its uncensored counterpart. So if I want to have the same method work on the 35B model, do I also need to install the official Qwen 35B model too?

EDIT: Nevermind! Turns out I edited the lines wrong lol 😅 Works fine for me now!

[heby1]

I actually stumbled on a situation where Thinking mode just activated by itself after one prompt, after having been off for the duration of the current chat. It was just normal story-like text content.

I added (NOTE TO ASSISTANT: Do not use Thinking Mode.) to my prompt and it solved that problem. This happened when using my solution above without editing any files. Probably making those config files would have prevented this one-off incident altogether.

[gungnirayu]

just add “{%- set enable_thinking = false %}”at the top of jinja