KwaiKAT Team Introduces KAT-Coder-V2.5: An Agentic Coding Model Trained on 100,000+ Verifiable Repositories
Anthropic
The KwaiKAT Team at Kuaishou has released KAT-Coder-V2.5, a coding model trained on over 100,000 verifiable environments. The model tops PinchBench at 94.9 and ranks second on SWE-Bench Pro at 65.2. An open-weight variant, KAT-Coder-V2.5-Dev, is available on Hugging Face under Apache-2.0.
The KwaiKAT Team at Kuaishou introduced KAT-Coder-V2.5, an agentic coding model trained on over 100,000 verifiable repositories spanning 12 languages. The model uses a pipeline called AutoBuilder that constructs executable environments with a success rate of 57.2%. Training infrastructure issues were fixed after an audit revealed ~16% of RL trajectories failed due to sandbox errors, which were reduced to below 2%. The model employs asymmetric PPO with a three-tier reward system and multi-teacher distillation. Under a unified Claude Code harness, KAT-Coder-V2.5 leads PinchBench with 94.9, beating Opus 4.8 at 93.5, and ranks second on SWE-Bench Pro (65.2 vs 69.2) and KAT Code Bench (53.1 vs 57.3), but lags on Terminal-Bench 2.1 (60.7) and SciCode (50.3). The open-weight KAT-Coder-V2.5-Dev is a separate 35B-total/3B-active MoE post-trained on Qwen3.6-35B-A3B, released on Hugging Face under Apache-2.0.
- Сокращения
- PPO = Proximal Policy Optimization — Проксимальная оптимизация политики
- GAE = Generalized Advantage Estimation — Обобщенная оценка преимущества
- MoE = Mixture of Experts — Смесь экспертов
- RL = Reinforcement Learning — Обучение с подкреплением
- SFT = Supervised Fine-Tuning — Контролируемая донастройка
Source: MarkTechPost —
original