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.
Figures & tables
Figure 1: Results of KAT-Coder-V2 and Claude Opus 4.6 across different scaffolds on various software engineering benchmarks.
Figure 2: KwaiEnv Workflow for SWE Tasks, Supporting Key Processes Including Data Synthesis, RL, and Evaluation
Expert
Scenario
Key Methodology
SWE
Issue resolution
Issue-PR pairing with merge-status supervision; AutoBuilder for verifiable task synthesis (F2P+P2P); Code Comprehension trajectory synthesis
Table 1: Overview of the five expert domains in the SFT stage.
Figure 3: Overview of the Issue-PR data construction pipeline. Merged PRs serve as anchor points for bidirectional Issue–PR mapping and code diff extraction, with merge status providing natural correctness supervision. The reconstructed engineering chains are decomposed into retrieval and editing tasks, then filtered through multi-stage quality control to yield over 2M training samples.
Figure 4: Overview of the AutoBuilder pipeline for automated SWE task synthesis from open-source repositories.
Figure 5: Overview of Tree Training in agentic RL.
Figure 6: Overview of the RL and sandbox framework.
Benchmark
Scaffold
KAT-Coder-V2
Claude Opus 4.6
SWE-bench Verified
Claude Code
79.6
80.8 ∗
OpenCode
74.8
75.0
OpenClaw
72.8
75.7
SWE-bench Multilingual
Claude Code
75.4
77.8 ∗
OpenCode
71.2
70.2
SWE-rebench-V2 (subset)
Claude Code
43.3
43.7
Table 2: Evaluation Results on Software Engineering Tasks under Multiple Scaffolds. Data points marked with * are taken from https://www.anthropic.com/news/claude-opus-4-6. Other data points are evaluated on KwaiEnv.
Benchmark
Scaffold
KAT-Coder-V2
GLM-5
MiniMax
Claude
GPT-5.4
Gemini
M2.7
Opus 4.6
3.1 Pro
PinchBench
Best Score
88.7
86.4
87.1
87.4
90.5
86.7
Average Score
81.9
80.3
81.8
82.3
81.6
75.9
Claw-Eval
Pass ˆ 3
55.6
57.7
51.9
66.3
66.3
50.0
Average Score
73.4
73.0
70.7
79.3
80.6
74.2
Table 3: Evaluation Results on Real-World Agent Benchmarks under OpenClaw. Data points of other models are taken from https://pinchbench.com/ and https://claw-eval.github.io/ (retrieved on March 25, 2026).
Benchmark
KAT-Coder-V2
GLM-5
Kimi K2.5
Landing Page
59.8
57.6
54.6
Slides
57.6
42.8
34.8
Data Visualization
67.6
42.4
46.0
Table 4: Evaluation Results on Frontend Generation Task.
Figure 7: Frontend Generation Results of KAT-Coder-V2.
Benchmark
KAT-Coder-V2
GLM-5
MiniMax
Claude
GPT-5.4
Gemini
M2.7
Opus 4.6
3.1 Pro
Terminal-Bench Hard
46.8
43.2
39.4
46.2
57.6
53.8
τ2 -Bench Telecom
93.9
98.2
84.8
92.1
91.5
95.6
AA-LCR
68.0
63.3
68.7
70.7
74.0
72.7
IFBench
67.0
72.3
75.7
53.1
73.9
77.1
Table 5: Evaluation Results on General Task. Data points of KAT-Coder-V2 are evaluated on KwaiEnv. Data points of other models are from https://artificialanalysis.ai/evaluations (retrieved on March 25, 2026).