Kimi新模型拿下代码开源SOTA，仅仅72B，发布即开源

Core Viewpoint - Kimi has released a new open-source model, Kimi-Dev, which achieved a state-of-the-art (SOTA) score of 60.4% on SWE-bench Verified, demonstrating strong programming capabilities with only 72 billion parameters [1][2]. Group 1: Model Performance and Comparison - Kimi-Dev outperforms the latest DeepSeek-R1 and shows excellent performance compared to closed-source models [2]. - Some observers believe that the capabilities of Kimi-Dev may have been underestimated, suggesting it could surpass xAI [3]. Group 2: Model Architecture and Training - Kimi-Dev integrates two roles: BugFixer and TestWriter, following a minimal framework with two phases [6]. - The model was trained using approximately 150 billion high-quality real data, including millions of GitHub issues and PR submissions, to enhance its prior knowledge [8]. - The training process involved strict data purification to exclude content from SWE-bench Verified [10]. Group 3: Training Methodology - The reinforcement learning phase utilized a strategy optimization method from Kimi k1.5, focusing on outcome-based rewards, efficient prompt sets, and positive example reinforcement [11][13][14]. - Kimi-Dev can generate up to 40 patch candidates and 40 test candidates for each problem during self-play testing, showcasing a scaling effect [16]. Group 4: Future Plans - The Kimi team plans to explore more complex software engineering tasks and deeper integration with popular IDEs, version control systems, and CI/CD pipelines [18].