Skip to main content

Kuaishou's New AI Model Can Build Entire Software Projects on Its Own

Kuaishou's KwaiKAT team has just dropped a bombshell in the AI coding world. Their latest model, KAT-Coder-Pro V2.5, isn't just another autocomplete tool—it's an AI agent that can take on entire software projects from start to finish. Think of it as a junior developer that never sleeps, never complains, and actually gets things done.

From Code Snippets to Full Projects

Most coding assistants today are great at filling in a line or two, but they stumble when faced with a real-world codebase. KAT-Coder-Pro V2.5 aims to change that. The team built a custom pipeline called AutoBuilder that can set up runnable repository environments with a success rate of 57.2%—a huge leap from the industry average of 16.5%. It supports 12 programming languages and has been tested on over 100,000 real repositories.

But here's the kicker: the model doesn't just write code. It can navigate across files, follow project conventions, and even debug itself when things go wrong. The team fed it high-quality training data from failed attempts, turning mistakes into learning opportunities.

A Toolbox That Keeps Growing

KwaiKAT also introduced KwaiClawEnv, a dynamic system that expands the model's toolkit on the fly. It generates complex workflows from real business tasks—things like data analysis, integrating different systems, or processing batches of documents. The system filters out low-quality trajectories, keeping only the best examples for training. The result? The model can handle tasks with more than 10 steps without breaking a sweat.

Training That Breaks the Mold

The team ditched the usual supervised fine-tuning approach in favor of large-scale reinforcement learning. They trained the model across multiple agent frameworks to avoid it getting stuck in one interaction style. An asymmetric PPO architecture solves the credit assignment problem—basically, figuring out which actions led to success in long, multi-step tasks. A hierarchical reward system balances core task results, behavioral norms, and exploration incentives.

To top it off, they used MOPD (Multi-teacher Online Policy Distillation) to merge the capabilities of five expert models: long-term engineering, general agentic skills, terminal usage, front-end aesthetics, and general knowledge. So a single model can now write code, run workflows, and generate front-end pages without switching contexts.

Benchmarks That Speak Volumes

The numbers are impressive: on SWE-Bench Pro, a standard coding benchmark, the model scored 65.2. On their internal KAT Code Bench, it hit 53.1—handling complete issues without manual decomposition. For agentic tasks, PinchBench gave it 94.2, and their own KAT Claw Bench scored 85.5. These aren't just incremental improvements; they represent a genuine shift in what AI can do in software development.

What This Means for Developers

KAT-Coder-Pro V2.5 is now available on the StreamLake platform (streamlake.com), with API access and technical documentation. The team has also released technical reports and set up developer discussion groups. For now, it's a powerful tool for automating repetitive coding tasks, but the long-term vision is clear: AI that can manage entire software projects, from concept to deployment.

Image

Image

Key Points

  • KAT-Coder-Pro V2.5 can handle complete software engineering tasks end-to-end, not just code completion.
  • AutoBuilder pipeline boosts repository environment setup success rate from 16.5% to 57.2%.
  • KwaiClawEnv dynamically expands tool capabilities for complex multi-step workflows.
  • Uses large-scale reinforcement learning with asymmetric PPO and hierarchical rewards.
  • MOPD distillation merges five expert models into one versatile agent.
  • Available now on StreamLake platform with API access.