What is Kimi-Dev
Kimi-Dev is an open-source code model launched by Moonshot AI, designed specifically for software engineering tasks. The model has 72B parameters, and its programming level is stronger than the latestDeepSeek-R1, and it also performs excellently compared with closed-source models. It achieves 60.4% performance on the SWE-bench Verified dataset, surpassing other open-source models and becoming the current SOTA among open-source models. Kimi-Dev is based on reinforcement learning and self-play mechanisms, and can efficiently fix code errors and write test code. The model is open-sourced under the MIT license, and the model weights, code, and related resources can be found on Hugging Face and GitHub.
Main features of Kimi-Dev
- Code repair (BugFixer): Automatically locates errors in code and generates repair patches, resolving vulnerabilities and defects in software development.
- Test code generation (TestWriter): Writes unit tests for existing code to ensure code quality and functional correctness.
- Automated development workflow: Based on reinforcement learning and self-play mechanisms, coordinates repair and testing capabilities to improve development efficiency.
- Integration with development tools: In the future, it will be deeply integrated with IDEs, version control systems, and CI/CD pipelines, seamlessly fitting into development workflows.
Technical principles of Kimi-Dev
- Role division (BugFixer and TestWriter): The model plays two roles, respectively responsible for repairing code and writing tests, both sharing the framework for file localization and code editing.
- Mid-training: Trained with approximately 150 billion high-quality data to enhance the model's understanding of actual development tasks.
- Reinforcement Learning: Runs code in a Docker environment and gives rewards based on test results, improving the model's code editing capabilities.
- Test-time Self-Play: During the testing phase, the model generates multiple patches and test candidates, and based on the self-play mechanism coordinates repair and testing capabilities to improve overall performance.
Performance of Kimi-Dev
- Performance on the SWE-bench Verified dataset：
- Comparison with open-source models: Kimi-Dev-72B achieves 60.4% performance on the SWE-bench Verified dataset, surpassing all other open-source models and becoming the current SOTA (State of the Art) among open-source models.
- Comparison with closed-source models: As a model with 72B parameters, its performance is already close to and in some aspects surpasses some closed-source models, showing strong competitiveness.
Project address of Kimi-Dev
- Project official website：https://moonshotai.github.io/Kimi-Dev/
- GitHub repository：https://github.com/MoonshotAI/Kimi-Dev
- HuggingFace model library：https://huggingface.co/moonshotai/Kimi-Dev-72B
Application scenarios of Kimi-Dev
- Code repair: Quickly detects and repairs errors or vulnerabilities in code, reducing the time for manual troubleshooting and repair.
- Test code generation: Generates high-quality unit test code for existing code, improving test coverage.
- Development workflow automation: Integrates with mainstream IDEs, providing real-time code repair suggestions and test code generation features.
- Programming education: Helps beginners quickly understand and master programming skills, assisting learning through generated example code and test code.
- Open-source project maintenance: Help maintainers of open source projects quickly fix vulnerabilities and optimize code, improving project quality and stability.