Released by Kuaishou, KAT-Coder V2.5 is a 35B-parameter coding model derived from the Alibaba Qwen3.6-35B-A3B base. The model was fine-tuned using 127,000 supervised examples and a specialized reinforcement learning (RL) pipeline called KuaiPilot. To build a robust training dataset, developers utilized "AutoBuilder," a dual-agent system that increased repository reconstruction rates from 16.5% to 57.2% across 12 programming languages. This enabled the creation of over 100,000 verifiable environments. To improve training accuracy, the team addressed a 16% error rate in the original sandbox and synchronized exploration and training tokens to prevent drift. Following 10 RL epochs with specific penalties for pathologies like excessive tool calls or self-repetition, mislabeled tool calls dropped from 9.34% to 0.28%. The model achieved significant benchmark scores, including 69.4 on SWE-bench Verified, 63 on a multilingual set, and 45.96 on SWE-bench Pro. While it trails frontier models like Claude Opus 5, its 21GB quantized size allows it to run on consumer hardware, such as an RTX 3060, demonstrating superior efficiency in reasoning and tool-calling compared to its base model.