Kuaishou's KAT-Coder-V2.5 Takes Entire GitHub Issues Autonomously — Open Weights Included
Kuaishou's AI team, KwaiKAT, released KAT-Coder-V2.5 in early July — an agentic coding model built around a different premise than most coding assistants. Rather than completing code snippets or answering questions about functions, KAT-Coder-V2.5 is designed to accept a whole GitHub issue or business task as input and work through it autonomously: navigating the repository, locating relevant context across files, implementing the changes, and running its own verification steps before returning a result.
The model was trained using an end-to-end agentic post-training framework the KwaiKAT team calls AutoBuilder, which reconstructs multilingual repositories into sandboxed environments to generate training data at scale. A second component, KwaiClawEnv, synthesizes large-scale tool-use trajectories — so rather than learning code from static examples, the model learned to act inside real, executable codebases. The team published a technical report alongside the release with benchmark results on PinchBench and repository-level software engineering tasks, where it claims state-of-the-art performance.
KAT-Coder-V2.5 supports a 256,000-token context window, which gives it room to hold multiple source files, dependency trees, and documentation simultaneously — important for the kind of cross-file reasoning autonomous issue resolution actually requires. The model also retains a multi-expert architecture from the V2 generation that lets it switch between logic-heavy backend work and front-end tasks without a separate model call.
Two versions shipped. KAT-Coder-Pro V2.5 is the commercial release, available via StreamLake and API. KAT-Coder-V2.5-Dev is an open-weight Mixture-of-Experts model — 35B total parameters, 3B activated per forward pass — published on Hugging Face for community research and evaluation. The open-weight release puts it in a similar position to GLM-5.2 in terms of inspectability: researchers can study what the training process actually produced, not just what the benchmark numbers say.
KAT-Coder-V2.5 arrives in the same window as several other agentic coding pushes — it's a different bet from models that target raw benchmark performance and instead focuses on whether the model can actually work inside a real repository without hand-holding.
