Kwaipilot · 2026-07-26 · major
KAT-Coder-V2.5-Dev — Kwaipilot's 35B open-weight agentic coding MoE
Kwaipilot dropped KAT-Coder-V2.5-Dev on Hugging Face under Apache 2.0. The 35B/3B-active MoE, built on Qwen3.6-35B-A3B, scores 69.4% on SWE-bench Verified and targets real repo work.
Kwaipilot opens the weights of a 35B/3B-active MoE tuned to act inside real code repositories, not just autocomplete a snippet.
Key specs
| SWE-bench | 69.4% |
|---|---|
| Params (total) | 35B |
| Params (active) | 3B |
Quick facts
| Maker | Kwaipilot (Kuaishou) |
|---|---|
| License | Apache-2.0 |
| Architecture | MoE, 35B total / 3B active |
| Base model | Qwen3.6-35B-A3B |
| Training data | 100,000+ verifiable repo environments, 12 languages |
| Availability | Open weights on Hugging Face; API via StreamLake for Pro variant |
| Paper | arXiv:2607.05471 (KAT-Coder-V2.5 Technical Report) |
What is it?
KAT-Coder-V2.5-Dev is an open-weight agentic coding model from Kwaipilot, the AI4SE team at Kuaishou. The Dev variant is a mixture-of-experts model with 35B total and 3B active parameters, published on Hugging Face under Apache 2.0. It is the open sibling of the closed KAT-Coder-V2.5-Pro that Kwaipilot serves through StreamLake.
How does it work?
Built on Qwen3.6-35B-A3B, the model is post-trained with the KAT-V2.5 recipe: supervised fine-tuning plus reinforcement learning with asymmetric PPO and a three-tiered reward. Kwaipilot's AutoBuilder tool reconstructs 100,000+ executable repositories across 12 languages, and KwaiClawEnv synthesizes multi-step tool-use trajectories the agent then trains against.
Why does it matter?
Open-weight coding models have crowded the mid-tier of SWE-bench Verified, and KAT-Coder-V2.5-Dev's 69.4% score puts it near the top of that pack without an API bill. Teams that need an agent to open files, run tests, and iterate in a real repo can now do so on their own hardware under an Apache 2.0 license.
Who is it for?
coding-agent builders, open-source infra teams, and self-hosters who want a permissively-licensed 35B MoE for repo-level automation
Frequently asked questions
- Is KAT-Coder-V2.5-Dev free to use?
- KAT-Coder-V2.5-Dev is released under Apache 2.0 on Hugging Face, so anyone can download the weights, run them locally, and use the model commercially without paying Kwaipilot. The stronger closed KAT-Coder-V2.5-Pro variant is API-only and served via Kuaishou's StreamLake product.
- How does KAT-Coder-V2.5-Dev compare to the Pro version?
- KAT-Coder-V2.5-Dev is the open-weight sibling of KAT-Coder-V2.5-Pro. The Dev variant scores 69.4% on SWE-bench Verified and 41.02% on Terminal-Bench 2.1, while the closed Pro variant tops PinchBench at 94.9 and ranks second on SWE-Bench Pro behind Claude Opus 4.8.
- What base model was KAT-Coder-V2.5-Dev built on?
- KAT-Coder-V2.5-Dev is post-trained on Qwen3.6-35B-A3B, Alibaba's mixture-of-experts model with 35B total and 3B active parameters. Kwaipilot kept the base architecture and applied the KAT-V2.5 SFT plus reinforcement-learning recipe used for the closed Pro model.
- What is KAT-Coder-V2.5-Dev trained to do?
- KAT-Coder-V2.5-Dev is an agentic coding model designed to work inside real, executable repositories instead of emitting single-turn snippets. Kwaipilot trained it on more than 100,000 verifiable repo environments across 12 languages using AutoBuilder for setup and asymmetric PPO for reinforcement learning.
Try it
huggingface-cli download Kwaipilot/KAT-Coder-V2.5-Dev