KAT-Coder V2.5

KAT-Coder V2.5 is an open-weight agentic coding model from Kwaipilot, the developer-tools group inside Kuaishou/Kwai. The Dev release landed on Hugging Face on July 23, 2026, pairing a Qwen3.6-35B-A3B MoE base (35B total, 3B active per token) with Apache 2.0 weights. Under a unified Claude Code evaluation harness it posts the best agentic tool-use score on PinchBench and ranks second on SWE-Bench Pro and KAT Code Bench, behind only the frontier Opus 4.8 and ahead of the GLM-5 series and Kimi-K2.6. The checkpoint has drawn roughly 582 likes and 23K downloads.

Model Specifications

Key Features

• Agentic tool use: Trained to act autonomously inside real executable repositories, verified by fail-to-pass and pass-to-pass tests.
• Repo-level engineering: AutoBuilder reconstructs real repositories into training tasks, teaching project-scale edits rather than single-file snippets.
• Verifiable RL training: KwaiClawEnv synthesizes tool-use trajectories, and hindsight-augmented value estimation stabilizes the PPO loop.
• OpenAI-compatible serving: Serve through SGLang, vLLM, or the Hugging Face Transformers server and point existing agent clients at the endpoint.
• Long context: 262K native tokens cover large monorepos and long agent sessions.

Use Cases

• Autonomous coding agents: Multi-step fixes where the model runs tests, reads failures, and iterates.
• Self-hosted assistants: Near-frontier agentic coding for teams that cannot send source to closed APIs.
• Research baseline: Apache 2.0 weights plus the arXiv report (2607.05471) suit agentic RL research.
• Cost-sensitive production: 3B active parameters keep inference far cheaper than dense frontier models.

Pricing

The Dev weights are free under Apache 2.0; you pay only for compute, and community GGUF quants from bartowski and mudler lower the bar further. A hosted commercial API runs on Kwaipilot's StreamLake platform at streamlake.com/product/kat-coder with usage-based pricing.

Advantages vs Competitors

1. vs GLM-5 and Kimi-K2.6: Beats both on SWE-Bench Pro and KAT Code Bench under the same harness, and leads PinchBench tool use outright.
2. vs closed Claude Code models: Only Opus 4.8 scores higher; KAT-Coder's open weights and self-hosting freedom are categorically different.
3. Efficiency: 3B active per token is a fraction of GLM-5 class activation, compounding at scale.

Getting Started

Serve with vLLM 0.19.0+. The language-model-only flag is required because Dev ships text weights only:

vllm serve Kwaipilot/KAT-Coder-V2.5-Dev --language-model-only

Or with SGLang 0.5.10+:

python -m sglang.launch_server \
  --model-path Kwaipilot/KAT-Coder-V2.5-Dev \
  --tp-size 8 --reasoning-parser qwen3

Both expose an OpenAI-compatible API; keep the qwen3 reasoning parser enabled.

FAQ

Does KAT-Coder V2.5 support vision inputs?

No. The base supports image-text-to-text, but Dev ships language-model weights only.

Can I use it commercially?

Yes. Apache 2.0 permits commercial use, modification, and redistribution.

How does it compare to Claude Code's models?

Second only to Opus 4.8 on SWE-Bench Pro and KAT Code Bench, and the open-weight leader on PinchBench.

Alternatives

• Qwen3 8-27B: the broader Qwen3 ecosystem KAT-Coder builds on.
• GLM-5.3: Zhipu's flagship coding model with a strong agentic story.
• Claude Code: the closed product defining the frontier this model chases.

Tips

1. Keep the reasoning parser on: Without --reasoning-parser qwen3 in SGLang, reasoning traces leak into answers.
2. Verify in your own harness: Replay real fail-to-pass suites before migrating.
3. Evaluate on a GGUF quant first: Smoke-test on smaller GPUs before an 8-way tensor-parallel server.

Conclusion

KAT-Coder V2.5 is the strongest open-weight agentic coding release of the season: a 35B MoE activating just 3B per token, trained with verifiable RL in real repositories, and benchmarked second only to Opus 4.8 under a Claude Code harness. For near-frontier agentic coding with Apache 2.0 freedom, this is the checkpoint to download.