Moonshot AI has unveiled Kimi-Dev-72B, a powerful open-source coding large language model (LLM) focused on software engineering and automated issue resolution. With a reported 60.4% performance on the SWE-bench Verified benchmark, Kimi-Dev-72B now leads among open-source models, setting a new state-of-the-art and outperforming previous contenders on practical software engineering tasks.
Kimi-Dev-72B distinguishes itself by leveraging large-scale reinforcement learning. The model is trained to autonomously patch real-world code repositories inside Docker containers, rewarding itself only when the complete test suite for the relevant software passes. This unique approach ensures that Kimi-Dev-72B´s code solutions are not only syntactically correct but also functionally robust, aligning closely with professional software development standards. The deployment and learning setup mimics end-to-end development workflows, making the model´s outputs directly applicable to real-world programming challenges.
This model is available to the wider community through both Hugging Face and GitHub, inviting developers, researchers, and organizations to use, test, and contribute further improvements. The quick start guide showcases how to integrate and interact with the model in Python using the widely adopted transformers library. The Kimi-Dev team also highlights their commitment to open science, with a forthcoming technical report and active community contributions encouraged. Released under the permissive MIT license, Kimi-Dev-72B underscores a collaborative approach to advancing coding-centric artificial intelligence tools and benchmarks.
