KwaiKAT Team Releases KAT-Coder-V2.5
The KwaiKAT Team at Kuaishou has introduced KAT-Coder-V2.5, a coding model trained to operate inside real, executable repositories rather than emit single-turn code. The served model is available through StreamLake. An open-weight variant, KAT-Coder-V2.5-Dev, was released separately on Hugging Face under Apache-2.0.
AutoBuilder and verifiable environments
The research frames a verifiable task as a triplet: a precise task description, an executable repository environment, and validation tests. A patch is correct only if it passes all tests. Tasks are mined from real pull requests and commits in the SWE-bench lineage. AutoBuilder reconstructs multilingual repositories into sandboxed environments; combining a preconfigured base environment, build-system templates, and distilled build recipes raised construction success from 16.5% to 57.2%, yielding over 100,000 verifiable environments across 12 languages. Git history and exploitable traces are stripped so agents cannot read the reference solution.
Training and infrastructure
KwaiKAT treats agentic coding as an infrastructure problem. A sandbox audit found ~16% of RL trajectories failed because of sandbox issues rather than the policy; fixes cut that rate below 2%. Training uses asymmetric actor-critic PPO, harness randomization, and Multi-Teacher On-Policy Distillation.
Results
Under a unified Claude Code harness, KAT-Coder-V2.5 leads on PinchBench at 94.9 (ahead of Opus 4.8 at 93.5), places second on SWE-Bench Pro (65.2 vs Opus 4.8 at 69.2), and lags on Terminal-Bench 2.1 (60.7 vs Opus 4.8 at 84.6). The open-weight KAT-Coder-V2.5-Dev is a separate 35B-total / 3B-active MoE post-trained on Qwen3.6-35B-A3B; its results are not comparable to the flagship table.
Paper: https://arxiv.org/html/2607.05471
Product: https://streamlake.com/product/kat-coder