Kat Coder Pro

Kat Coder Pro is an agentic coding model from Kuaishou's Kwaipilot team, built for real-world software engineering tasks such as feature implementation, bug fixing, and refactoring. It pairs a 256K token context window with up to 128K output tokens, so it can work across large files and multi-step projects in a single session. The model reaches a 73.4% solve rate on SWE-Bench Verified, a benchmark that measures how often a model resolves real software issues. It was trained through a multi-stage pipeline that combines mid-training, supervised fine-tuning, reinforcement fine-tuning, and large-scale agentic reinforcement learning. Kat Coder Pro is tuned for tool use and multi-tool parallel invocation, multi-turn interaction, and instruction following across many programming scenarios. It is a text-only, non-reasoning model, returning direct responses without chain-of-thought overhead, which suits coding agents and automated engineering pipelines that need fast completions.