Kuaishou KwaiKAT Releases KAT-Coder-Pro V2.5: Say Goodbye to Code补 - The First Domestic Agentic Programming Model That Can Run the Entire Project End-to-End

KwaiKAT team officially released the flagship Agentic Coding model KAT-Coder-Pro V2.5 today. The new version systematically upgrades around three dimensions: long-term engineering capabilities, general Agentic capabilities, and large-scale Agentic reinforcement learning system, aiming to evolve from single code completion to an AI agent capable of independently completing complete software engineering tasks and complex business workflows.

Regarding long-term engineering capabilities, the team developed its own AutoBuilder automated pipeline, increasing the success rate of building runnable repository environments from an industry average of about 16.5% to 57.2%, covering 12 programming languages and over 100,000 verified real repository environments. At the same time, high-value failure trajectories are recycled into training data, enabling the model to master cross-file positioning, following project specifications, and self-debugging testing capabilities.

On the general Agentic capabilities, KwaiKAT has developed the KwaiClawEnv system — a dynamically expanding tool pool, deriving massive complex workflows from real business tasks, retaining high-quality training trajectories through dual filtering, covering scenarios such as data analysis, cross-system integration, and batch document processing, supporting task execution of more than 10 links.

In the training aspect, the team abandoned the pure supervised fine-tuning approach and adopted large-scale Agentic reinforcement learning: by using Harness Scaling, training was conducted across various mainstream Agent frameworks to avoid overfitting to a single interaction format; introducing asymmetric PPO architecture to solve the credit assignment problem in long-term tasks; designing a hierarchical reward mechanism (core task results + behavioral norm constraints + failure exploration incentives) to balance effectiveness and robustness. The model further uses MOPD multi-teacher online strategy distillation to integrate the capabilities of five expert models: long-term engineering, general Agentic, terminal usage, front-end aesthetics, and general knowledge. A single model can now simultaneously handle multiple scenarios such as writing code, running workflows, and generating front-end pages, without switching.

Official evaluation data shows: in code engineering, SWE-Bench Pro score is 65.2, and internal KAT Code Bench score is 53.1, directly handling complete Issues without manual decomposition; in Agentic tasks, PinchBench score is 94.2, and internal KAT Claw Bench score is 85.5, with excellent full-process stability.

Currently, KAT-Coder-Pro V2.5 has been fully launched on StreamLake platform (streamlake.com), open for API application and technical documentation access, and also publicly releasing technical reports and developer exchange groups.

Address: https://streamlake.com/product/kat-coder