KAT-Coder-Pro & Air

KAT-Coder is a flagship agentic coding model developed by Kuaishou designed for end-to-end software engineering tasks such as bug fixing, feature implementation, and repository-level refactoring. By leveraging large-scale agentic reinforcement learning, it acts as an autonomous developer capable of navigating codebases, executing tests, and iterating on solutions.

KAT-Coder is a flagship agentic coding model developed by Kuaishou designed for end-to-end software engineering tasks such as bug fixing, feature implementation, and repository-level refactoring. By leveraging large-scale agentic reinforcement learning, it acts as an autonomous developer capable of navigating codebases, executing tests, and iterating on solutions.

• 01Autonomous agentic workflows for multi-step software engineering tasks
• 02Multi-file reasoning and repository-level code navigation
• 03Compatible with 10+ mainstream AI coding scaffolds (e.g., Claude Code, Cline, OpenHands)
• 04Training optimized via multi-stage reinforcement learning (RFT/RL)
• 05Supports parallel tool calling and long-context interaction