KAT-Coder-V2: Agentic Coding Model Hits 79.6% SWE-bench via Specialize-then-Unify Training

arXiv·high signal

Kuaishou's KAT-Coder-V2 decomposes agentic coding into five expert domains (SWE, WebCoding, Terminal, WebSearch, General), trains each independently with SFT and RL, then consolidates via on-policy distillation into a single model. Achieves 79.6% on SWE-bench Verified (vs Claude Opus 4.6 at 80.8%) and 88.7 on PinchBench surpassing GLM-5 and MiniMax M2.7. Built on KwaiEnv infrastructure sustaining tens of thousands of concurrent sandbox instances, with a novel Tree Training technique achieving 6.2x speedup for tree-structured RL trajectories.