KAT-Coder-V2 Technical Report

2026-03-29

AI Summary

This work addresses the challenge of balancing domain-specific expertise and general-purpose capability in multi-domain code generation by proposing a “specialize-then-unify” architecture. The approach first decomposes programming tasks into five expert domains, each refined through supervised fine-tuning and reinforcement learning, and subsequently unifies them into a single model via online policy distillation. Key innovations include the MCLA algorithm to stabilize MoE-based reinforcement learning, Tree Training to accelerate computation over tree-structured execution trajectories, and KwaiEnv—a modular sandbox environment supporting tens of thousands of concurrent evaluations. The resulting model achieves state-of-the-art performance across multiple benchmarks: 79.6% on SWE-bench Verified, 88.7 on PinchBench, top rankings in three frontend aesthetics metrics, 46.8 on Terminal-Bench Hard, and 93.9 on tau²-Bench.

Abstract

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.

Problem

• agentic coding
• code generation
• multi-domain programming
• intelligent software engineering
• generalist coding model

Innovation

• Specialize-then-Unify
• KwaiEnv
• MCLA
• Tree Training
• MoE RL