Kimi-Dev-72B: An Open-Source Coding LLM for Issue Resolution

Kimi-Dev-72B, developed by the Kimi-Dev Team, is a 72.7 billion parameter open-source language model specifically engineered for software engineering tasks. It distinguishes itself by achieving a new state-of-the-art performance among open-source models on the challenging SWE-bench Verified benchmark.

Key Capabilities and Differentiators

• State-of-the-Art Performance: Kimi-Dev-72B scores 60.4% on SWE-bench Verified, surpassing other open-source models in its ability to resolve real-world software issues.
• Reinforcement Learning Optimization: The model is optimized through large-scale reinforcement learning, where it autonomously patches real repositories within Docker environments. It receives rewards only upon successfully passing the entire test suite, ensuring high-quality and robust solutions.
• Real-World Alignment: This training methodology aligns Kimi-Dev-72B with practical development standards, making it highly effective for generating correct and verifiable code.

Ideal Use Cases

• Automated Bug Fixing: Excellent for tasks involving the identification and resolution of software bugs.
• Code Generation and Refactoring: Capable of generating and improving code that adheres to testing standards.
• Software Engineering Research: A valuable tool for researchers exploring advanced code intelligence and autonomous development.