Cerebras uses Kimi K3 (3 trillion parameters) to show how to split a model across 59 WSE wafers: 1,000–2,000 tokens per second claimed
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
The largest open models are starting to answer at speeds close to live conversation, if Cerebras's own data is to be believed.
In a Cerebras blog post published on October 22, the company explains how it places trillion-parameter-scale models: layers are spread across several wafers linked by RDMA over Ethernet. The open Kimi K3 (about 3 trillion parameters, 69 linear-attention layers and 24 dense ones) is used as an illustration: for it the post gives 59 wafers, a speed of 1,000–2,000 or more tokens per second and about 73 concurrent requests with a 256,000-token context. The figures are the company's own.
Sources 1