Kwaipilot releases KAT-Coder-V2.5-Dev, a 35B MoE model targeting agentic coding

Kwaipilot/KAT-Coder-V2.5-Dev · Hugging Face

Kwaipilot open-sourced KAT-Coder-V2.5-Dev on Hugging Face: a 35B MoE with 3B active params, tuned via SFT and RL for agentic coding. They claim SOTA at this scale and cut abnormal tool-label rates from 9.34% to 0.28%. Reddit commenters note the Qwen 3.6 35B SWE-bench numbers in their table are much lower than the official model card, and suspect gains partly come from using Claude Code as the training harness. The post doesn't include other coding benchmarks.

Why it matters: KAT-Coder-V2.5-Dev delivers a concrete metric (abnormal tool-call rate 9.34% → 0.28%) on a 35B/3B MoE for agentic coding — H and K both hit. But the team is unknown, Reddit has one post, R is absent, and the post doesn't disclose benchmark baselines or RL details. Lands right ...