Kimi-Dev-72B is an open-source large language model specifically designed for software engineering tasks, particularly issue resolution and code repair. It targets developers and researchers seeking state-of-the-art performance in automated code correction and testing, offering a robust solution for improving software development workflows.

How It Works

Kimi-Dev-72B employs a two-stage framework: File Localization and Code Editing. It first identifies relevant files for modification based on issue descriptions and repository context, then performs precise code edits for bug fixes or test insertions. This file-level localization approach allows for more detailed reasoning in the subsequent editing stage, differentiating it from multi-step localization methods. The model is further optimized through large-scale reinforcement learning, where it autonomously patches real repositories in Docker, receiving rewards only upon successful test suite completion, ensuring robust and correct solutions.

Quick Start & Requirements

• Installation: Clone the repository, create a conda environment (conda create -n kimidev python=3.12), and install locally (pip install -e .).
• Prerequisites: CUDA 12.8 is required for vLLM installation (pip install vllm --extra-index-url https://download.pytorch.org/whl/cu128). Preprocessed SWE-bench data is available for download.
• Deployment: Serve the model using vLLM: serve Kimi-Dev-72B --served-model-name kimi-dev --host 0.0.0.0 --port 8000 --gpu-memory-utilization 0.95 --max-seq-len-to-capture 131072 --tensor-parallel-size 8.
• Resources: Requires significant GPU memory (e.g., 0.95 utilization) and tensor parallelism (size 8 suggested).
• Links: Huggingface, Demo.

Highlighted Details

• Achieves state-of-the-art performance (60.4%) on SWE-bench Verified among open-source models.
• Optimized via reinforcement learning with autonomous patching and test-suite validation.
• Supports both bug fixing and test writing through its rollout scripts.
• Handles long contexts with max-seq-len-to-capture up to 131072.

Maintenance & Community

The project is maintained by the Kimi-Dev Team. Contributions are welcomed via Pull Requests or Issues. Contact is available via GitHub issues or zhuhan@moonshot.cn.

Licensing & Compatibility

The repository is available on GitHub and Hugging Face. Licensing details for commercial use or closed-source linking are not explicitly stated in the README.

Limitations & Caveats

The README mentions a tech report is "Coming soon," suggesting some documentation may be incomplete. The model's performance is benchmarked on SWE-bench Verified, and real-world applicability may vary. CUDA 12.8 is a specific requirement for vLLM.