API Error: 400 "qwen3-coder:30b" does not support thinking

[mydeveloperplanet]

What happened?

I am using the following setup:

• Ollama on a remote server, version 0.13
• I use the following model: https://ollama.com/library/qwen3-coder:30b
• qwen 0.6.0

When I enter test in my project directory, the following error is returned.

✕ [API Error: 400 "qwen3-coder:30b" does not support thinking]

What did you expect to happen?

I would have expected a successful response to be returned.

When I use https://ollama.com/library/gpt-oss:20b , the model wants to invoke the test target of my Maven project. This is ok.

Client information

Client Information

Run qwen to enter the interactive CLI, then run the /about command.

$ qwen /about
About Qwen Code                                                                  │
│                                                                                  │
│ CLI Version                 0.6.0                                                │
│ Git Commit                  17eb20c1                                             │
│ Model                       gpt-oss:20b                                          │
│ Sandbox                     no sandbox                                           │
│ OS Platform                 linux                                                │
│ OS Arch                     x64                                                  │
│ OS Release                  6.8.0-90-generic                                     │
│ Node.js Version             v24.12.0                                             │
│ NPM Version                 11.6.2                                               │
│ Session ID                  7e95e785-9c4f-4e23-8c49-e50eb84d4552                 │
│ Auth Method                 openai                                               │
│ Base URL                    https://<remote url>/v1                │
│ Memory Usage                281.5 MB    

Login information

API Config

Anything else we need to know?

I am a newbie using Qwen Coder, want to try it out in order to see whether we can use it

[AlxV07]

Hey @mydeveloperplanet

I had the same issue w/ pretty much the same setup (qwen3-coder:30b model + ollama 0.13.5 (locally running though) + qwen 0.6.0)

After doing some experimenting, it seems that qwen-code API requests always try to use the target model with think mode.
Even though the qwen3-coder models aren't thinking models, for some reason the option is still passed in the API request.

Funnily, this works when using ollama's qwen3-coder:480b-cloud in qwen-code: even though ollama warns that thinking isn't supported, the chat request still goes through and successfully generates. You can see this behavior for yourself when using ollama run {model}:

~ ollama run qwen3-coder:480b-cloud
Connecting to 'qwen3-coder:480b' on 'ollama.com' ⚡
>>> /set think 
warning: model "qwen3-coder:480b-cloud" does not support thinking output
Set 'think' mode.
>>> hi
Hello! How can I help you today?

>>> Send a message (/? for help)

And funnily, this doesn't work for qwen3-coder:30b:

~ ollama run qwen3-coder:30b       
>>> /set think
warning: model "qwen3-coder:30b" does not support thinking output
Set 'think' mode.
>>> hi
error: 400 Bad Request: "qwen3-coder:30b" does not support thinking
>>> Send a message (/? for help)

Which is the error we were getting.

Workaround solution:

If you want to use a non-thinking model, I found this workaround solution which has worked for me:

1. start a local server which serves as a proxy to ollama, intercepting and modifying chat requests (in our case, removing the thinking flag from the req)
2. change the base url in qwen-code from ollama to our proxy server
3. this may result in some weird context-length issues where prompts are truncated to 4096 tokens bc the context length options were ignored somehow (happened to me, haven't figured out why yet), but you can fix this by creating a new ollama model using ollama create {new_model_name} -f {modelfile} with the modelfile:

FROM qwen3-coder:30b
PARAMETER num_ctx 262144  # qwen3-coder max context length

Then you can set the model in qwen-code to your new model, and this works fully.

If you want, I can send my proxy server script (one Python file, uvicorn + FastAPI server)

[mydeveloperplanet]

Thanks for your reply and for the workaround. But still I do think that this should be solved in qwen-code. How ironically it is: qwen-code does not work with local qwen models.

[Christopher-R-Perkins]

You need to add a config to tell it not to use reasoning in the model object of your settings.json:

    "generationConfig": {
      "reasoning": false
    }

something like:

{
  "model": {
    "name": "qwen3-coder:30b",
    "generationConfig": {
      "reasoning": false
    }
  }
}

Qwen code defaults to a medium reasoning level when this config isn't here(you can technically pass it the reasoning level here as well).

[mydeveloperplanet]

Thanks @Christopher-R-Perkins , this works!

The generationConfig parameters are limited documented...

[mydeveloperplanet]

Correction @Christopher-R-Perkins , I tested with the wrong model, when I test this with qwen3-coder, still the same error is returned. However, i do notice that @tanzhenxin has fixed this , so I do hope this will be released quickly

[tanzhenxin]

@mydeveloperplanet It has been fixed in release v0.6.1. Please try it out and let me know if it works for you. Thx!

[mydeveloperplanet]

It works, thanks!