🤖Moonshot's Kimi K2.7-Code Cuts Thinking Tokens by 30%
TL;DR
Moonshot AI released Kimi K2.7-Code, an open-weight coding model on the same trillion-parameter MoE as K2.6, claiming about 30% lower thinking-token use. It ships on Hugging Face under a modified MIT license, but practitioners say its in-house benchmarks don't hold up.
Moonshot AI released Kimi K2.7-Code, an open-weight coding model on the same trillion-parameter MoE as K2.6, claiming about 30% lower thinking-token use. It ships on Hugging Face under a modified MIT license, but practitioners say its in-house benchmarks don't hold up.
Key Points
Open-weight release this week on Hugging Face with an OpenAI-compatible API
Trillion-parameter mixture-of-experts built on K2.6, with roughly 30% fewer reasoning tokens
Vendor claims +21.8% on Kimi Code Bench v2, +11% on Program Bench, +31.5% on MLS Bench Lite, all proprietary
Weights and code released under a modified MIT license
Independent testers question whether the benchmark gains reproduce
Why It Matters
China's open-weight coding models keep undercutting closed US labs on cost, but self-run benchmarks make real capability gains hard to verify.
Quick Facts
Frequently Asked Questions
Why does this matter?
China's open-weight coding models keep undercutting closed US labs on cost, but self-run benchmarks make real capability gains hard to verify.
What happened?
Moonshot AI released Kimi K2.7-Code, an open-weight coding model on the same trillion-parameter MoE as K2.6, claiming about 30% lower thinking-token use. It ships on Hugging Face under a modified MIT license, but practitioners say its in-house benchmarks don't hold up.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 2,042 builders reading daily.