KAT-Coder-Pro V1 is a specialized large language model developed by KwaiKAT, the AI for Software Engineering (AI4SE) division of Kuaishou. Designed specifically for agentic coding tasks, the model is optimized for real-world software engineering scenarios, including tool usage, multi-turn interaction, and complex logical reasoning. It features a massive context window of 256,000 tokens, enabling it to process and reason over extensive codebases or project documentation in a single session.
The model utilizes a Mixture-of-Experts (MoE) architecture, reported to feature approximately 72 billion active parameters within a larger total parameter framework. It gained prominence for its performance on technical benchmarks, notably achieving a 73.4% solve rate on the SWE-Bench Verified leaderboard. Its development involves a multi-stage training pipeline comprising mid-training for knowledge injection, supervised fine-tuning (SFT), reinforcement fine-tuning (RFT), and scalable agentic reinforcement learning (RL).
KAT-Coder-Pro V1 is primarily accessible as a flagship proprietary model through the StreamLake platform. While the Pro version remains closed-source, KwaiKAT has released complementary open-weights variants under the "KAT-Dev" series for research and development purposes.