Kuaishou’s KwaiKAT Team Releases KAT-Coder-V2.5, an Open-Weight Agentic Coding Model Trained on 100,000+ Verifiable Repositories

The KwaiKAT Team at Chinese tech company Kuaishou has released KAT-Coder-V2.5, an agentic coding model that operates inside real, executable software repositories rather than generating single-turn code snippets. The model and its accompanying technical report argue that agentic coding capability is bottlenecked by training infrastructure rather than model scale.

KAT-Coder-V2.5 is available through Kuaishou’s StreamLake platform, while an open-weight variant, KAT-Coder-V2.5-Dev, has been released separately on Hugging Face under the Apache-2.0 license. This makes the model accessible to Indian developers and startups building AI-powered coding tools.

AutoBuilder: Solving the Environment Construction Bottleneck

One of the core innovations in KAT-Coder-V2.5 is AutoBuilder, a system that automatically constructs verifiable repository environments. The team found that environment construction success rates were initially only 16.5%. After combining preconfigured base environments, build-system templates, and a retrievable library of distilled build recipes, AutoBuilder raised the success rate to 57.2%. This produced over 100,000 verifiable environments spanning 12 programming languages.

Each verifiable task in the training pipeline is defined as a triplet: a precise task description, an executable repository environment, and a set of validation tests. Tasks are mined from real pull requests and commits following the SWE-bench methodology. The merged code change provides a golden patch, and the accompanying test change provides a test patch. Raw issue text is discarded and instead regenerated into three components: a problem statement grounded in the golden patch, requirements derived from the test patch, and interface constraints inferred from both. A clarity check removes any ambiguous, incomplete, or internally inconsistent descriptions.

Infrastructure Audits Cut Reward Noise

During training of the earlier KAT-Coder-V2, slow reward curves led the team to investigate the sandbox infrastructure. An audit revealed that about 16% of trajectories failed due to sandbox infrastructure issues rather than the model policy. Boundary misalignments sometimes emptied observations for roughly 40 steps and corrupted rewards.

Three infrastructure fixes reduced the sandbox feedback error rate from approximately 16% to below 2%. First, an early-release image eviction policy lowered disk usage from 95% to 60%, cutting timeout-induced invalid rollouts from 6–7% to under 1%. Second, correcting environment variables during remote sandbox initialization stopped system overrides that flipped rewards on 6–7% of samples. Third, the Gateway Server bypassed mainstream chat endpoints, which caused 40% token drift at around 200-turn scales, and instead called the generate endpoint directly to ensure rollout token alignment. Training collapses dropped by an order of magnitude.

Training with Asymmetric Actor-Critic and Multi-Teacher Distillation

The team used Proximal Policy Optimisation (PPO) with Generalised Advantage Estimation (GAE), employing an asymmetric actor–critic setup. The Critic receives privileged training context (rewards, tests, coverage, patches, metadata, future turns), while the Actor sees only rollout state. At inference, the Critic and extra context are discarded.

Rewards are structured into three tiers: core task scores (requiring all fail-to-pass and pass-to-pass tests to pass), standard behaviour constraints (penalising duplication, bad tool calls, and debug remnants), and failed-trajectory incentives (scoring file retrieval and giving partial test credit). Five expert models are fused via Multi-Teacher On-Policy Distillation using reverse KL divergence, an off-policy start, and drift-aware truncation.

Benchmark Performance

Under a unified Claude Code harness, KAT-Coder-V2.5 leads its panel on PinchBench with a score of 94.9, beating Opus 4.8 at 93.5. It places second on SWE-Bench Pro (65.2 versus 69.2) and on the internal KAT Code Bench (53.1 versus 57.3). However, it lags on Terminal-Bench 2.1, placing last with 60.7 behind GLM-5.1 (61.8) and Opus 4.8 (84.6). On SciCode, it scores 50.3, matching GLM-5.2.

The open-weight KAT-Coder-V2.5-Dev is a separate 35B-total / 3B-active Mixture-of-Experts model post-trained on Qwen3.6-35B-A3B using 127K supervised fine-tuning examples, then reinforcement learning. Its benchmark results are not directly comparable to the main flagship table.

Why This Matters for Indian Developers

The release of an open-weight agentic coding model under Apache-2.0 gives Indian AI startups and developer tool companies access to a model trained on real, verifiable repository environments. The infrastructure innovations — particularly AutoBuilder and sandbox reliability fixes — could be adopted by local teams building similar training pipelines. The model is available for download and experimentation on Hugging Face.

Datos clave:

Aspect	Detail
Model release	KAT-Coder-V2.5 (flagship) and KAT-Coder-V2.5-Dev (open-weight, Apache-2.0)
Training data	Over 100,000 verifiable environments across 12 languages using AutoBuilder
Key infrastructure fix	Sandbox feedback error rate reduced from ~16% to <2%
Top benchmark score	9 on PinchBench, surpassing Opus 4.8

Source: MarkTechPost – “KwaiKAT Team Releases KAT-Coder-V2.5: An Agentic Coding Model Trained on 100,000+ Verifiable Repository Environments” (https://www.marktechpost.com/2026/07/26/kwaikat-team-releases-kat-coder-v2-5-an-agentic-coding-model-trained-on-100000-verifiable-repository-environments/)