Single-Node Agentic Inference Hits 1,000 tok/s

JiaZhihao · x · 2026-07-16

Lithos reported massive interest in their ultra-low latency agentic inference preview, attracting 100+ companies and thousands of visitors within days. As a result, they've expanded capacity and extended the public preview to next weekend.

They claim that this agentic inference engine serves Kimi K2.7 Code on a single standard 8xB200 node, peaking at over 1,000 tok/s/用户 on coding workloads.

Original post →

More from coding & agent

coding & agent channel →