Kwai releases 35B MoE coding model as open weights

·2026.07.23 20:31

Key point

Kwai released KAT-Coder-V2.5-Dev, an open-weight coding model with a 35B MoE architecture.

Details

Kwai's AI coding team Kwaipilot released KAT-Coder-V2.5-Dev on Hugging Face. It's a MoE (Mixture of Experts) architecture with 35B total parameters and 3B activated parameters.

Key features:

Achieved Agentic Coding SOTA among models of similar parameter scale through SFT and RL training
Abnormal tool label occurrence rate reduced from 9.34% to 0.28% (-9pp)
Single-turn repetition errors eliminated from 0.34% to 0%

This is the open-weight version of KAT-Coder-V2.5, which launched in July, released for the purpose of sharing research with the community.

llm huggingface moe agentic-coding coding-model open-weights kwaipilot

Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.