Skip to main content

Kuaishou's New AI Model Writes Code, Not Just Completes It

Kuaishou's KwaiKAT team has just dropped a new version of their AI coding model, KAT-Coder-Pro V2.5, and it's not your average autocomplete. This one aims to do the whole job—from reading a bug report to writing the fix, running tests, and making sure everything works. Think of it as an AI that can handle a software project end-to-end, not just fill in a few lines.

What's New Under the Hood?

The team focused on three big areas: long-term engineering skills, general agentic abilities, and a massive reinforcement learning system. The result? A model that can tackle complex tasks that span multiple files and require understanding the entire project structure.

Building a Better Environment

One of the biggest headaches for AI coding tools is setting up a working development environment. The team built their own automated pipeline called AutoBuilder, which boosts the success rate of creating runnable environments from an industry average of about 16.5% to a whopping 57.2%. It supports 12 programming languages and has verified over 100,000 real repository setups. That means the AI can actually run and test the code it writes.

Learning from Failure

Instead of just training on successful code, the team also feeds the model high-value failure trajectories—cases where the code didn't work. This helps the model learn to debug itself, follow project conventions, and navigate across files. It's like learning from your mistakes, but at scale.

General Agentic Capabilities

The model isn't just for coding. KwaiKAT developed the KwaiClawEnv system, a dynamic tool pool that generates complex workflows from real business tasks. It covers data analysis, cross-system integration, batch document processing, and more. The system uses dual filtering to keep only the highest-quality training trajectories, ensuring the model can handle tasks with over 10 steps.

Training: Ditching Supervised Fine-Tuning

The team moved away from pure supervised fine-tuning and adopted large-scale agentic reinforcement learning. They used a technique called Harness Scaling to train across multiple agent frameworks, avoiding overfitting to a single interaction style. An asymmetric PPO architecture helps solve the credit assignment problem in long tasks—figuring out which actions led to success or failure. A hierarchical reward mechanism balances core task results, behavioral norms, and failure exploration incentives.

Combining Expert Models

The model also 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. So a single model can now write code, run workflows, and generate front-end pages without switching between different tools.

Performance Numbers

On the SWE-Bench Pro benchmark, the model scored 65.2, and on the internal KAT Code Bench, it scored 53.1. These tests involve handling complete issues without manual decomposition. For agentic tasks, it scored 94.2 on PinchBench and 85.5 on the internal KAT Claw Bench, showing strong stability across full processes.

Availability

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.

Image

Image

Key Points

  • KAT-Coder-Pro V2.5 is an agentic coding model that handles entire software projects end-to-end.
  • AutoBuilder pipeline increases environment setup success rate from 16.5% to 57.2%.
  • Trained with large-scale reinforcement learning, not just supervised fine-tuning.
  • Integrates five expert models for versatile capabilities.
  • Available on StreamLake platform with API access.