Models

A lightweight 35B-A3B MoE variant of KAT-Coder-V2.5, built for long-horizon agentic coding.

Agentic coding: 69.4% on SWE-bench Verified and 93.4% on PinchBench.
Long-context: native support for up to 1M tokens for longer-context serving.
Reliable tool use: fewer malformed tool labels and repetitive single-turn behaviors.

KAT-Coder-V2.5

KAT-Coder-V2.5 Technical Report

KwaiKAT Team, Kuaishou Technology

KAT-Coder-V2.5 achieves 65.2% on SWE-bench Pro (vs. Claude Opus 4.8 at 69.2%), 94.2% on PinchBench (vs. Claude Opus 4.8 at 93.5%)
KAT-Coder-V2.5 surpassing GLM-5.1, approaching GLM-5.2

KAT-Coder-V2

KAT-Coder-V2 Technical Report

KwaiKAT Team, Kuaishou Technology

KAT-Coder-V2 ranks 6th on Artificial Analysis Coding Index
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)

LPM

LPM: Industrial-Scale Generative Video Restoration

LPM Team, Kuaishou Technology

Large Processing Model (LPM) is a diffusion-based generative framework for photorealistic video restoration
To our knowledge, LPM is the first generative video restoration model deployed at industrial scale

KAT-Coder-V2.5-Dev