The KwaiKAT Team at Kuaishou has released KAT-Coder-V2.5, a coding model designed to operate within real, executable repository environments rather than generating single-turn code. The team also published an open-weight variant, KAT-Coder-V2.5-Dev, on Hugging Face under the Apache-2.0 license.
- AutoBuilder constructs verifiable task triplets (description, environment, tests) from real pull requests, raising environment construction success rates from 16.5% to 57.2%.
- The resulting dataset includes over 100,000 verifiable environments spanning 12 languages, with git history and commit metadata stripped to prevent solution leakage.
- A data scaling flywheel filters trajectories by process quality rather than just final test success, using targeted hints for near-misses and rule-based gates for passing runs.
- Infrastructure fixes reduced sandbox feedback errors from ~16% to below 2%, addressing issues like disk usage timeouts and token drift in the Gateway Server.
- Training utilizes asymmetric PPO with a three-tier reward system and Multi-Teacher On-Policy Distillation, discarding privileged critic context at inference.
KAT-Coder-V2.5 leads PinchBench with a score of 94.9 and ranks second on SWE-Bench Pro with 65.2, demonstrating improved performance in agentic coding tasks.