Cover Image

KAT-Coder-V2 Technical Report

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Li, Fengxiang, Zhang, Han, Huang, Haoyang, Wang, Jinghui, Hao, Jinhua, Yuan, Kun, Li, Mengtong, Zhang, Minglei, Xu, Pengcheng, Zhuang, Wenhao, Shao, Yizhen, Feng, Zongxian, Tang, Can, Wang, Chao, Tong, Chengxiao, Yang, Fan, Xiong, Gang, Gao, Haixuan, Gao, Han, Wang, Hao, Liu, Haochen, Sun, Hongliang, Li, Jiabao, Chang, Jingwen, Du, Jun, Peng, Junyi, Cui, Leizhen, Jing, Meimei, Wu, Mingqi, Yan, Shangpeng, Qi, Shaotong, Xu, Suzhe, Zhao, Wenxuan, Sun, Xianda, Xie, Xuan, Wang, Yanbo, Xia, Yao, Cui, Yinghan, Chen, Yingpeng, Wang, Yong, Shi, Yuze, Shen, Zhiwei, Wang, Ziyu, Sun, Ming, Ye, Lin, Chen, Bin
Format:	Preprint
Published:	2026
Subjects:	Computation and Language Machine Learning
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

_version_	1866911551326257152
author	Li, Fengxiang Zhang, Han Huang, Haoyang Wang, Jinghui Hao, Jinhua Yuan, Kun Li, Mengtong Zhang, Minglei Xu, Pengcheng Zhuang, Wenhao Shao, Yizhen Feng, Zongxian Tang, Can Wang, Chao Tong, Chengxiao Yang, Fan Xiong, Gang Gao, Haixuan Gao, Han Wang, Hao Liu, Haochen Sun, Hongliang Li, Jiabao Chang, Jingwen Du, Jun Peng, Junyi Cui, Leizhen Jing, Meimei Wu, Mingqi Yan, Shangpeng Qi, Shaotong Xu, Suzhe Zhao, Wenxuan Sun, Xianda Xie, Xuan Wang, Yanbo Xia, Yao Cui, Yinghan Chen, Yingpeng Wang, Yong Shi, Yuze Shen, Zhiwei Wang, Ziyu Sun, Ming Ye, Lin Chen, Bin
author_facet	Li, Fengxiang Zhang, Han Huang, Haoyang Wang, Jinghui Hao, Jinhua Yuan, Kun Li, Mengtong Zhang, Minglei Xu, Pengcheng Zhuang, Wenhao Shao, Yizhen Feng, Zongxian Tang, Can Wang, Chao Tong, Chengxiao Yang, Fan Xiong, Gang Gao, Haixuan Gao, Han Wang, Hao Liu, Haochen Sun, Hongliang Li, Jiabao Chang, Jingwen Du, Jun Peng, Junyi Cui, Leizhen Jing, Meimei Wu, Mingqi Yan, Shangpeng Qi, Shaotong Xu, Suzhe Zhao, Wenxuan Sun, Xianda Xie, Xuan Wang, Yanbo Xia, Yao Cui, Yinghan Chen, Yingpeng Wang, Yong Shi, Yuze Shen, Zhiwei Wang, Ziyu Sun, Ming Ye, Lin Chen, Bin
contents	We present KAT-Coder-V2, an agentic coding model developed by the KwaiKAT team at Kuaishou. KAT-Coder-V2 adopts a "Specialize-then-Unify" paradigm that decomposes agentic coding into five expert domains - SWE, WebCoding, Terminal, WebSearch, and General - each undergoing independent supervised fine-tuning and reinforcement learning, before being consolidated into a single model via on-policy distillation. We develop KwaiEnv, a modular infrastructure sustaining tens of thousands of concurrent sandbox instances, and scale RL training along task complexity, intent alignment, and scaffold generalization. We further propose MCLA for stabilizing MoE RL training and Tree Training for eliminating redundant computation over tree-structured trajectories with up to 6.2x speedup. KAT-Coder-V2 achieves 79.6% on SWE-bench Verified (vs. Claude Opus 4.6 at 80.8%), 88.7 on PinchBench (surpassing GLM-5 and MiniMax M2.7), ranks first across all three frontend aesthetics scenarios, and maintains strong generalist scores on Terminal-Bench Hard (46.8) and tau^2-Bench (93.9). Our model is publicly available at https://streamlake.com/product/kat-coder.
format	Preprint
id	arxiv_https___arxiv_org_abs_2603_27703
institution	arXiv
publishDate	2026
record_format	arxiv
spellingShingle	KAT-Coder-V2 Technical Report Li, Fengxiang Zhang, Han Huang, Haoyang Wang, Jinghui Hao, Jinhua Yuan, Kun Li, Mengtong Zhang, Minglei Xu, Pengcheng Zhuang, Wenhao Shao, Yizhen Feng, Zongxian Tang, Can Wang, Chao Tong, Chengxiao Yang, Fan Xiong, Gang Gao, Haixuan Gao, Han Wang, Hao Liu, Haochen Sun, Hongliang Li, Jiabao Chang, Jingwen Du, Jun Peng, Junyi Cui, Leizhen Jing, Meimei Wu, Mingqi Yan, Shangpeng Qi, Shaotong Xu, Suzhe Zhao, Wenxuan Sun, Xianda Xie, Xuan Wang, Yanbo Xia, Yao Cui, Yinghan Chen, Yingpeng Wang, Yong Shi, Yuze Shen, Zhiwei Wang, Ziyu Sun, Ming Ye, Lin Chen, Bin Computation and Language Machine Learning We present KAT-Coder-V2, an agentic coding model developed by the KwaiKAT team at Kuaishou. KAT-Coder-V2 adopts a "Specialize-then-Unify" paradigm that decomposes agentic coding into five expert domains - SWE, WebCoding, Terminal, WebSearch, and General - each undergoing independent supervised fine-tuning and reinforcement learning, before being consolidated into a single model via on-policy distillation. We develop KwaiEnv, a modular infrastructure sustaining tens of thousands of concurrent sandbox instances, and scale RL training along task complexity, intent alignment, and scaffold generalization. We further propose MCLA for stabilizing MoE RL training and Tree Training for eliminating redundant computation over tree-structured trajectories with up to 6.2x speedup. KAT-Coder-V2 achieves 79.6% on SWE-bench Verified (vs. Claude Opus 4.6 at 80.8%), 88.7 on PinchBench (surpassing GLM-5 and MiniMax M2.7), ranks first across all three frontend aesthetics scenarios, and maintains strong generalist scores on Terminal-Bench Hard (46.8) and tau^2-Bench (93.9). Our model is publicly available at https://streamlake.com/product/kat-coder.
title	KAT-Coder-V2 Technical Report
topic	Computation and Language Machine Learning
url	https://arxiv.org/abs/2603.27703