KAT-Coder: Kuaishou’s Coding Agent That Learns Inside Real Software Projects

Kuaishou’s KAT-Coder is a coding agent trained to work inside real, executable software repositories instead of memorizing code text. The open-weight KAT-Coder-V2.5-Dev scores 69.4 on SWE-bench Verified, and the flagship model reaches 73.4%.
artificial-intelligence
software-engineering
Author

Kabui, Charles

Published

2026-08-26

Keywords

kat-coder, kuaishou, agentic-coding, open-weights, swe-bench

Download as PDF

KAT-Coder is Kuaishou’s coding agent: an AI that works a task end to end, editing files, running commands, and calling tools until the job is done. What sets it apart is how it learns. The Kwaipilot team’s AutoBuilder pipeline rebuilds open-source and enterprise projects into sandboxed environments with real tests, then runs large-scale reinforcement learning, so the model practices on actual codebases instead of memorizing code text. KAT-Coder-V2.5-Dev, the open-weight version released in late July 2026, is a 35-billion-parameter model that uses only 3 billion at a time, under the permissive Apache-2.0 license, and it scores 69.4 on SWE-bench Verified, a standard test of fixing real bugs in open-source projects. The flagship KAT-Coder reaches 73.4% and, per the V2.5 technical report, ranks second only to Claude Opus 4.8 on repository-level software engineering.

Both KAT-Dev-32B and V2.5-Dev are downloadable, so a developer can run them locally or wire them into tools like Claude Code through a small proxy instead of renting one vendor’s agent. Kuaishou reports two emergent behaviors from the reinforcement learning stage: the model finishes tasks in 32% fewer interaction turns and can call several tools at once, which shows up as faster, cheaper agent runs. The strongest KAT-Coder is free to try on the company’s StreamLake platform.

Training coding models inside executable environments with verifiable tests, rather than on static code, is where the field’s biggest gains are now coming from. KAT-Coder is a clear example of learning by doing replacing learning by reading.

Read More: OpenAI Codex CLI: a Rust-built open-source coding agent

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {KAT-Coder: {Kuaishou’s} {Coding} {Agent} {That} {Learns}
    {Inside} {Real} {Software} {Projects}},
  date = {2026-08-26},
  url = {https://toknow.ai/posts/kat-coder-kuaishou-agentic-coding-model/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “KAT-Coder: Kuaishou’s Coding Agent That Learns Inside Real Software Projects.” https://toknow.ai/posts/kat-coder-kuaishou-agentic-coding-model/.