Run qwen3 coder locally

[simkimsia]

2025-08-31 Sunday UTC

I finally got moatless tools to run for models that run locally, in this case Qwen3-Coder 30b via LM Studio.

Ollama was not suitable. I used swebench_react for the --flow. I also had to change code in 3 places for this to work.

I don't expect this to be merged without changes, so I'm happy to listen to feedback for this to be suitable to be merged.

I only manage to do the Verify Setup step in README. I haven't run a full evaluation. But it took me a long time to even get to this point, so I will return to make run_evaluation.py work with Qwen3-Coder on LM Studio

[aorwall]

Forgot to reply to your discussion post, sorry for that. But good job getting it to work with the non-existent documentation 😅

Model config will read from the file models.json in the directory you set in the env var MOATLESS_DIR. So it might work to put the following in that file:

  {
    "model_id": "qwen-3-coder",
    "model": "openai/qwen3-coder-30b",
    "temperature": 0.0,
    "max_tokens": 8000,
    "timeout": 120.0,
    "model_base_url": "http://host.docker.internal:1234/v1",
    "model_api_key": "faux_key",
    "metadata": null,
    "message_cache": true,
    "thoughts_in_action": false,
    "disable_thoughts": false,
    "few_shot_examples": true,
    "headers": {},
    "params": {},
    "merge_same_role_messages": false,
    "reasoning_effort": null,
    "completion_model_class": "moatless.completion.tool_call.ToolCallCompletionModel"
  }

And then run uv run python scripts/docker_run.py --flow swebench_tools --model-id qwen-3-coder --instance-id django__django-11099 --evaluation-name testing_setup