Open source code generation model for software engineering tasks

By /

this project Kimi‑Dev It is an open source “Coding LLM” released by Moonshot AI, designed to automatically repair code and write tests. Its latest model is Kimi‑Dev‑72B(approximately 72.7 billion parameters), achieved on the SWE‑bench Verified benchmark in the open source field 60.4% has achieved the best open source model performance at present

🚀Core highlights

  • Twin frames: BugFixer + TestWriter
    • BugFixer Locate error files and repair code
    • TestWriter Write unit tests and generate test cases that fail and pass
      The two work together to ensure complete and verifiable repairs
  • Mid-term Training + Intensive Learning (RL)
    • Use nearly 150 billion tokens of training data from real GitHub issues/PR to enhance model code understanding and positioning capabilities
    • Using the RL strategy of “reward for passing a test”, the model actually runs tests in a Docker environment, and rewards will be awarded only if all tests pass.
  • Self-confrontation mechanism during testing (Self-Play)
    • BugFixer and TestWriter “play self-game” against each other during the inference phase, generating multiple candidate patches and tests, and ultimately selecting the best result

📦Usage

  1. Clone warehouse and install environment:git clone https://github.com/MoonshotAI/Kimi-Dev.git conda create -n kimidev python=3.12 pip install -e .
  2. Install vLLM (supporting CUDA 12.8) and start the inference service:pip install vllm --extra-index-url https://download.pytorch.org/whl/cu128 vllm serve Kimi‑Dev‑72B --served-model-name kimi-dev ...
  3. Run using the script provided by the project:
    • rollout_messages_bugfixer.py Used for automatic repairs
    • rollout_messages_testwriter.py for generating test
  4. Model weights and demos are available on Hugging Face

📈 Benchmark results and background

  • in SWE‑bench Verified stands at 60.4% The problem resolution rate is higher than that of other similar models.
  • The project code, models and usage documentation have been published on GitHub and Hugging Face, and are open source under the MIT license.
  • Moonshot AI is a China AI company established in 2023. It is positioned to develop high-performance, multi-modal, large-context LLM. It also has Kimi‑k1.5/Kimi‑VL/Kimi‑Audio and other series.

🧠Summary

Kimi‑Dev is a powerful open source programming LLM that focuses on automated code repair and unit test generation, and has achieved the best results in open source in related tasks. If your needs are:

  • Automatically fix project bugs
  • Generate or supplement tests for the project
  • Research the implementation application of AI in software engineering

Then this project is very worthy of in-depth study. You can further review its README, code structure, sample scripts (such as kimidev/ folder), or deploy a Demo to get started.

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

Oil tubing:

Scroll to Top