KAT-Coder-V2.5: Coding agent beats all but Opus 4.8 on repo-level SWE

Researchers from Kuaishou (alongside 52 co-authors) have unveiled KAT-Coder-V2.5, a coding-focused agentic model designed not for single-turn code generation but for autonomous operation inside real, executable repositories. The team identified that performance is bottlenecked less by model scale than by the scarcity of reproducible environments, verifiable rewards, and high-value trajectories. To overcome this, they built an end-to-end agentic post-training framework.

At its core are two infrastructure engines: AutoBuilder, which reconstructs multilingual repositories into sandboxed environments with fail-to-pass and pass-to-pass verification at scale, allowing the model to regenerate task specifications and recover near-miss trajectories via process-aware filtering; and KwaiClawEnv, which synthesizes large-scale tool-use trajectories from executable services and real task seeds. The training pipeline then scales reinforcement learning with harness randomization, a reliability-hardened sandbox, and an asymmetric actor-critic PPO with hindsight-augmented value estimation. A harness-oriented reward framework further guides the learning, and the final model unifies expertise from SWE, Agent-Claw, and WebCoding domains using multi-teacher on-policy distillation.

On evaluation, KAT-Coder-V2.5 delivers the best agentic tool-use result on the PinchBench benchmark and ranks second only to the frontier Opus 4.8 on repository-level software engineering tasks. The model is available via a service URL, and the technical report details 24 pages of methodology including 5 figures. This work demonstrates a practical path to advancing coding agents through environment simulation and scalable RL, bypassing the need for larger base models.