KAT-Coder

model paper
Kuaishou
72B active parameter MoE coding model using the "AutoThink" paradigm. Designed for agentic coding with parallel tool calls to reduce interaction latency. 73.4 on SWE-Bench Verified.
Paper (arXiv)Project PageAPI DocsArtificial Analysis (KAT-Coder-Pro V2)

Outputs 3

KAT-Coder

model

72B active parameter MoE coding model using the "AutoThink" paradigm. Designed for agentic coding with parallel tool calls to reduce interaction latency. Served via the StreamLake API as KAT-Coder-Pro V1.

API DocsProject PagePaper (arXiv)Artificial Analysis (KAT-Coder-Pro V1)
Architecture MOE
Active params 72B
AA Intelligence 19 was 22 on v4.2

KAT-Coder-Pro V2

model

Kwaipilot's 2026 agentic-coding refresh of the KAT-Coder line. <strong>AA Intelligence Index v4.3: 22</strong> (non-reasoning). Base undisclosed; the KAT line is post-trained from external pretrained bases (Kwaipilot's own pipeline description is "fine-tuned a pretrained model", the open V1 sibling KAT-Dev-72B-Exp is architecturally Qwen2.5-72B, and the open V2.5 sibling is built on Qwen3.6-35B-A3B).

Artificial AnalysisAPI Docs
Architecture MOE
AA Intelligence 22 was 26 on v4.2

KAT-Coder Technical Report

paper

73.4 on SWE-Bench Verified. Agentic coding model technical report.

Paper (arXiv)Project Page
arXiv HTML
codingmoeagenticreasoning