Kwaipilot/KAT-Dev

KAT-Dev (32B) is an open-source 32B-parameter model specifically designed for software engineering tasks. In the SWE-Bench Verified benchmark, it achieved a solution rate of 62.4%, ranking fifth among all open-source models of varying sizes. The model was optimized through multiple stages, including intermediate training, supervised fine-tuning (SFT), reinforcement fine-tuning (RFT), and large-scale agent reinforcement learning (RL). Built upon Qwen3-32B, its training process laid the foundation for subsequent fine-tuning and reinforcement learning stages by enhancing fundamental capabilities such as improved tool utilization, multi-turn interactions, and instruction following. During the fine-tuning stage, the model not only learned eight carefully curated task types and programming scenarios but also innovatively introduced a reinforcement fine-tuning (RFT) stage, guided by “teacher trajectories” annotated by human engineers. Finally, the agent reinforcement learning stage addressed scalability challenges through multi-level prefix caching, entropy-based trajectory pruning, and an efficient architecture.