AB
AB
AiBoss

All Categories

project

Kimi-Dev - Open source code model launched by Dark Side of the Moon

Kimi-Dev is an open-source model from Moonshot AI, designed specifically for software engineering tasks. The model has 72 bytes of parameters, boasts superior programming capabilities compared to the latest DeepSeek-R1, and performs exceptionally well compared to closed-source models. In S...

What is Kimi-Dev?

Kimi-Dev is an open-source model from Moonshot AI, designed specifically for software engineering tasks. With 72 bytes of parameters, it boasts superior programming capabilities compared to the latest DeepSeek-R1 and outperforms closed-source models. Achieving a 60.4% performance on the SWE-bench Verified dataset, it surpasses other open-source models, becoming the current state-of-the-art (SOTA) open-source model. Based on reinforcement learning and self-play mechanisms, Kimi-Dev efficiently fixes code errors and writes test code. The model is open-source under the MIT license; model weights, code, and related resources can be found on Hugging Face and GitHub.

Kimi-Dev's main functions

  • Code Fix (BugFixer)Automatically locates errors in the code and generates fix patches to resolve vulnerabilities and defects in software development.
  • Test code generation (TestWriter)Write unit tests for existing code to ensure code quality and functional correctness.
  • Automated development processBased on reinforcement learning and self-game mechanisms, it coordinates repair and testing capabilities to improve development efficiency.
  • Integration with development toolsIn the future, it will be deeply integrated with IDEs, version control systems, and CI/CD pipelines, seamlessly integrating into the development workflow.

Kimi-Dev's technical principles

  • Role division (BugFixer and TestWriter)The model plays two roles: fixing code and writing tests, respectively, and both share the framework for file location and code editing.
  • Mid-trainingThe model is trained using approximately 150 billion high-quality data points to enhance its understanding of real-world development tasks.
  • Reinforcement LearningRewards are given based on test results when running code in a Docker environment, thereby improving the model's code editing capabilities.
  • Test-time self-playDuring the testing phase, the model generates multiple patches and test candidates, coordinating repair and testing capabilities based on a self-game mechanism to improve overall performance.

Kimi-Dev's performance

  • Performance on the SWE-bench Verified dataset:
    • Comparison with open source modelsKimi-Dev-72B achieved a performance of 60.4% on the SWE-bench Verified dataset, surpassing all other open-source models and becoming the current state-of-the-art (SOTA) open-source model.
    • Comparison with closed-source modelsAs a model with parameter 72B, its performance is close to or even surpasses some closed-source models in some aspects, showing strong competitiveness.

Kimi-Dev's project address

Application scenarios of Kimi-Dev

  • Code FixQuickly detect and fix errors or vulnerabilities in the code, reducing the time spent on manual investigation and repair.
  • Test code generationGenerate high-quality unit test code for existing code, improving test coverage.
  • Development process automationIt integrates with mainstream IDEs, providing real-time code fix suggestions and test code generation capabilities.
  • Programming EducationIt helps beginners quickly understand and master programming skills by generating example and test code to aid learning.
  • Open source project maintenanceIt helps maintainers of open-source projects quickly fix vulnerabilities and optimize code, thereby improving project quality and stability.
Back