On June 17th, the dark side of the moon launched a new open-source code model for software engineering tasks - Kimi-Dev-72B.
This model achieved the world's highest level of open source model performance in SWE bench Verified programming benchmark testing, surpassing the newly released DeepSeeker R1 with only 72B of parameters and 671B of parameters.
This model is optimized through large-scale reinforcement learning, capable of autonomously patching real repositories in Docker, and only receives rewards when the entire test suite passes, ensuring the correctness and robustness of the solution.
The design philosophy and technical details of Kimi-Dev-72B include the combination of BugFixer and TestWriter, mid-term training, reinforcement learning, and self game during testing.
Among them, the complementary design of BugFixer and TestWriter laid the foundation for the model, making it perform well in fixing errors and writing tests.
Mid term training enhances the model's understanding of actual bug fixes and unit testing through approximately 150 billion high-quality real-world data.
The reinforcement learning stage focuses on improving their code editing skills, enabling them to excel in file localization and code editing.
Currently, Kimi-Dev-72B is available for download and deployment on Hugging Face and GitHub, including model weights, source code, and more.
