開放模型／程式代理
KAT-Coder-V2.5-Dev Opens 35B MoE Weights, Runs Coding Agents with 3B Active Parameters
Kwaipilot has released the KAT-Coder-V2.5-Dev weights, fine-tuning Qwen3.6-35B-A3B on 127,000 samples before applying reinforcement learning on agentic tasks. The team reports 69.40% on SWE-bench Verified in its own evaluation, although the score is highly sensitive to the agent framework and tool configuration.
Kwaipilot released KAT-Coder-V2.5-Dev on July 24. It is a text-only coding agent model with a mixture-of-experts architecture, comprising 35 billion total parameters and activating approximately 3 billion parameters per token. The model card provides weights in Hugging Face Transformers format and deployment instructions for Transformers, vLLM, SGLang, and KTransformers. Multimodal components are not included in this release.
Built on Qwen3.6-35B-A3B, the model first underwent supervised fine-tuning on 127,000 samples, followed by reinforcement learning. The accompanying technical report describes the broader KAT-V2.5 training system: AutoBuilder reconstructs real-world repositories as executable sandboxes and creates verifiable rewards using fail-to-pass and pass-to-pass tests, while KwaiClawEnv synthesizes tool-use trajectories from services and real-task seeds. Training also incorporates agent-framework randomization to reduce the risk of the model overfitting to a single harness.
Using its internally built unified pipeline, the team reports scores of 69.40% on SWE-bench Verified, 63.00% on SWE-bench Multilingual, and 45.96% on SWE-bench Pro, as well as 41.02% on Terminal-Bench 2.1. The tool-tag anomaly rate also fell from 9.34% to 0.28%. However, the release team reran all of these comparisons, generally evaluating each model only once. The model card also acknowledges that Qwen3.6 scored roughly 10 percentage points lower on SWE-bench in its environment than the officially reported figure, potentially because of differences in harness versions and test-set processing.
Engineering teams therefore should not interpret leaderboard gaps directly as differences in model capability. A more meaningful test is whether the model retains an advantage in patch success rate, retry cost, and erroneous tool-call rate when evaluated with the same agent framework, container image, toolset, and token budget. It will also be important to see whether the community can reproduce the reported results and whether the open weights are sufficient to replicate the official post-training results.