An open-source large language model (LLM) named Kimi-Dev-72B has been released, aiming to revolutionize the way problems are solved in the software engineering field. The model demonstrates state-of-the-art performance on the SWE-bench Verified benchmark, excelling particularly in autonomously repairing real-world code repositories within Docker environments.

QQ20250617-095538.png

The core advantage of Kimi-Dev-72B lies in its optimization through large-scale reinforcement learning. It can independently identify and fix code defects and is rewarded when the entire test suite passes, ensuring that the provided solutions are not only correct but also robust.

The project is open-source on both Hugging Face and GitHub, allowing developers to freely download and deploy it. Kimi-Dev-72B provides a simplified two-stage framework specifically designed for code repair and test writing tasks, including precise file localization and efficient code editing functions. This makes it a powerful tool for developers to troubleshoot and maintain their code.

Address: https://github.com/MoonshotAI/Kimi-Dev