Kuaishou's KAT-Coder-V2.5 tops PinchBench with open Dev variant
Kuaishou's KwaiKAT team says its new agentic coder, trained by RL in 100,000+ verifiable repository environments, edges Claude Opus 4.8 on PinchBench.
Kuaishou's KwaiKAT team released KAT-Coder-V2.5 on July 26, an agentic coding model whose defining feature is its training substrate: reinforcement learning across more than 100,000 verifiable repository environments spanning 12 programming languages. The team combined PPO with an asymmetric actor-critic design, multi-teacher distillation and a three-tier reward structure, treating agentic coding less as a modeling problem than as an infrastructure problem — building executable repositories at scale so that every training signal can be checked against real code.
A jagged but credible scorecard
The model tops PinchBench at 94.9, edging out Claude Opus 4.8's 93.5, and lands second on SWE-Bench Pro at 65.2 against Opus 4.8's 69.2. It matches GLM-5.2 on SciCode at 50.3 but trails its peer group on Terminal-Bench 2.1 with 60.7 — an unusually candid profile in a market where labs typically publish only their winning numbers.
An open executor for local agent stacks
Alongside the flagship, Kuaishou published KAT-Coder-V2.5-Dev on Hugging Face under Apache-2.0: a mixture-of-experts model with 35 billion total and just 3 billion active parameters, post-trained from Qwen3.6-35B-A3B. That footprint is small enough for local serving, and it slots naturally into the planner–worker pattern that coding-agent products are converging on, where a cheap open model handles execution under a frontier planner.
The release matters for two reasons. First, verifiable repository-scale RL environments are emerging as the decisive input to coding-agent quality — the labs that can build and grade the most realistic environments improve fastest, and Kuaishou just demonstrated that capacity at the hundred-thousand scale. Second, it extends a pattern: Chinese consumer-internet companies — Kuaishou is best known abroad for its Kling video model — are now fielding credible coding models alongside dedicated AI labs, deepening an already crowded and fast-moving Chinese agentic-coding field.