Kimi-Dev – An open-source code model launched by Moonshot AI

What is Kimi-Dev?

Kimi-Dev is an open-source code model developed by Moonshot AI, specifically designed for software engineering tasks. With 72 billion parameters, the model surpasses the latest DeepSeek-R1 in coding capabilities and demonstrates competitive performance even when compared to closed-source models. On the SWE-bench Verified benchmark, Kimi-Dev achieves 60.4% accuracy, outperforming all other open-source models and establishing itself as the current state-of-the-art (SOTA) in the open-source domain.

Built on reinforcement learning and a self-play mechanism, Kimi-Dev can efficiently fix bugs and generate test code. The model is open-sourced under the MIT license, and its weights, source code, and related resources are available on Hugging Face and GitHub.

Key Features of Kimi-Dev

• Bug Fixing (BugFixer): Automatically identifies and fixes bugs in code by generating repair patches, addressing vulnerabilities and defects in software development.

• Test Code Generation (TestWriter): Generates unit tests for existing code to ensure code quality and correctness.

• Automated Development Workflow: Leverages reinforcement learning and self-play to coordinate bug fixing and test generation, improving development efficiency.

• Developer Tool Integration: Future plans include deep integration with IDEs, version control systems, and CI/CD pipelines for seamless workflow adoption.

Technical Foundations of Kimi-Dev

• Role Division (BugFixer & TestWriter): The model takes on two distinct roles—bug fixing and test generation—while sharing a unified framework for file navigation and code editing.

• Mid-Training: Trained on approximately 150 billion high-quality data tokens, enhancing the model’s understanding of real-world development tasks.

• Reinforcement Learning: Executes code in a Docker environment and receives rewards based on test outcomes, improving its code editing capabilities.

• Test-time Self-Play: During inference, the model generates multiple patch and test candidates. A self-play mechanism coordinates bug fixing and test generation, optimizing overall performance.

Kimi-Dev Performance

• On SWE-bench Verified Benchmark:

  • Compared to Open-Source Models: Kimi-Dev-72B achieves 60.4% accuracy, outperforming all other open-source models and becoming the SOTA in its category.

  • Compared to Closed-Source Models: Despite being open-source, Kimi-Dev shows competitive or even superior performance compared to some closed-source models, showcasing strong capabilities.

Project Links for Kimi-Dev

• Project Website: https://moonshotai.github.io/Kimi-Dev/

• GitHub Repository: https://github.com/MoonshotAI/Kimi-Dev

• Hugging Face Model Hub: https://huggingface.co/moonshotai/Kimi-Dev-72B

Application Scenarios for Kimi-Dev

• Bug Fixing: Quickly detects and fixes bugs or vulnerabilities in code, significantly reducing the time spent on manual debugging.

• Test Code Generation: Produces high-quality unit tests for existing code, improving test coverage and code reliability.

• Automated Development Workflows: Integrates with popular IDEs to provide real-time bug fixing suggestions and test generation capabilities.

• Programming Education: Assists beginners in understanding and mastering programming through generated example and test code.

• Open Source Project Maintenance: Helps maintainers of open-source projects quickly fix bugs and optimize code, improving overall project quality and stability.