Skip to main content

Kuaishou's KAT-Coder-Pro V2.5: AI That Writes Entire Projects, Not Just Code Snippets

Kuaishou's KwaiKAT team has just dropped a new AI model that's more than just a code autocomplete. Meet KAT-Coder-Pro V2.5, an "agentic" programming model that can take on entire software projects from start to finish—no human hand-holding required.

What's New?

The update focuses on three big areas: long-term engineering chops, general agentic skills, and a fancy reinforcement learning system. In plain English? The model can now handle complex, multi-step tasks without breaking a sweat.

Long-Term Engineering: Building Real Projects

One of the biggest headaches for AI coders is setting up a working environment. Most models fail at this—industry success rates hover around 16.5%. But KwaiKAT built their own automated pipeline called AutoBuilder, which boosts that to 57.2%. It covers 12 programming languages and over 100,000 real repository environments. When the model fails, those failures get recycled into training data, teaching it to navigate cross-file dependencies, follow project specs, and debug itself.

Image

General Agentic Skills: Beyond Code

KwaiKAT also developed KwaiClawEnv, a dynamic tool pool that generates complex workflows from real business tasks. Think data analysis, cross-system integration, batch document processing—tasks that span more than 10 steps. The system filters out low-quality training trajectories, keeping only the good stuff.

Image

Training: Reinforcement Learning at Scale

Forget plain supervised fine-tuning. The team went all-in on large-scale agentic reinforcement learning. They used something called Harness Scaling to train across multiple agent frameworks, avoiding overfitting to one interaction style. An asymmetric PPO architecture solves the credit assignment problem—basically, figuring out which actions led to success in long tasks. A hierarchical reward mechanism balances core task results, behavioral norms, and failure exploration incentives.

To top it off, they used MOPD multi-teacher online strategy distillation to blend the skills of five expert models: long-term engineering, general agentic, terminal usage, front-end aesthetics, and general knowledge. The result? A single model that can write code, run workflows, and generate front-end pages without switching modes.

Performance Numbers

Official benchmarks look impressive: SWE-Bench Pro score of 65.2, internal KAT Code Bench score of 53.1—handling complete issues without manual decomposition. On agentic tasks, PinchBench score is 94.2, and internal KAT Claw Bench score is 85.5. That's solid full-process stability.

Availability

KAT-Coder-Pro V2.5 is now live on StreamLake platform (streamlake.com). Developers can apply for API access and check out the technical docs. The team also released technical reports and developer discussion groups.

Key Points

  • AutoBuilder boosts environment setup success from 16.5% to 57.2%
  • KwaiClawEnv handles complex workflows with over 10 steps
  • Reinforcement learning with asymmetric PPO and hierarchical rewards
  • MOPD distillation combines five expert models into one
  • Available now on StreamLake platform