Alibaba's Qwen team open-sourced enhanced weights for Qwen2.5-Coder-32B-Instruct tailored for long-horizon software engineering. Incorporating execution feedback and synthetic multi-turn debugging trajectories, the model reaches a 48.9% solve rate on SWE-bench Verified, setting a new open-weights record for the 32B class with 128k context support on single-GPU hardware.
Key Takeaways
- ✓Sets a 48.9% SWE-bench Verified score in the 32B parameter class.
- ✓Reinforced alignment with execution feedback from real-world repository testing.
- ✓Weights live on Hugging Face and ModelScope with instant vLLM and Ollama support.