It worked very well for me using qwen3 coder behind a litellm. Most other models just fail in weird ways though.