KwaiKAT Team Releases KAT-Coder-V2.5: An Agentic Coding Model Trained on 100,000+ Verifiable Repository Environments

Most coding-model papers talk about bigger context windows or smarter prompting. KwaiKAT's writeup on KAT-Coder-V2.5 reads differently: it's an infrastructure post dressed up as a model release. The team built AutoBuilder, a system that mines real pull requests and commits, strips out the giveaway git history, and forces a build agent to write installation and test scripts from a clean checkout. A separate verification agent then checks that over 90% of expected tests actually run and reproduce consistently across attempts — not just that some exit code came back green. That pushed environment construction success from a dismal 16.5% up to 57.2%, and the payoff is a library of more than 100,000 verifiable environments spanning 12 languages.

The more interesting admission comes from their reinforcement learning phase. During KAT-Coder-V2 training, reward curves were crawling, and the team initially assumed the RL algorithm itself was underperforming. An audit told a different story: roughly 16% of trajectories were failing because of sandbox plumbing, not because the model made bad decisions. Boundary misalignments were sometimes wiping observations for 40 steps straight, silently corrupting the reward signal the whole training run depended on. Three fixes — evicting stale disk images earlier, correcting environment variables during sandbox startup, and routing generation calls directly to /generate instead of through chat endpoints that were quietly re-tokenizing and drifting by 40% at long turn counts — dropped that error rate from 16% to under 2%. Training collapses fell by roughly tenfold as a result.

On the data side, KwaiKAT didn't just filter trajectories by pass/fail, which they found actively misleading: some passing runs were hard-coded shortcuts, some failing runs contained genuinely useful debugging behavior. So near-miss trajectories get targeted hints, which lift previously zero-pass tasks to roughly a 20% pass rate, and then the model is retrained hint-free so it doesn't cheat off information it wouldn't have at inference. Passing runs go through rule-based filtering and a multi-dimensional scoring pass covering things like patch minimality and honesty. They also randomize tool names, argument formats, and prompts across identical tasks to stop the model from memorizing harness quirks rather than actually solving problems.```

On training method, the team leaned on PPO with an asymmetric actor-critic setup — the critic gets privileged access to test results and future turns during training, then gets discarded entirely at inference, leaving just the actor. Five expert models get fused into the final system via on-policy distillation. Results-wise, KAT-Coder-V2.5 tops PinchBench at 94.9 against Opus 4.8's 93.5, but comes in second on SWE-Bench Pro and dead last on Terminal-Bench 2.1. Kuaishou also shipped a separate open-weight version, KAT-Coder-V2.5-Dev, a 35B-total/3B-active MoE built on Qwen3.6, released under Apache-2.0 — though its benchmark numbers come from a different evaluation protocol, so they're not directly comparable to the flagship's.