SettledHardware
Cerebras details serving a 3T-parameter model across 59 wafers at 1,000+ tokens a second
In an Oct. 22 engineering post, Cerebras describes running Kimi K3, an open model of about 3 trillion parameters, on 59 WSE-3T wafers with layers pipelined across them, and cites 1,000 to 2,000+ tokens per second. It assumes 4-bit weights and an 8-bit KV cache. The post gives no pricing and points to a free API.
- HEAT
- OUTLETS
- 1
- HACKER NEWS
- 348 pts
- FIRST SEEN
- LAST SIGNAL
Primary source
Cerebrascerebras.ai/blog/how-we-deploy-trillion-parameter-frontier-models-on-cerebrasCoverage · 1 article, oldest first
Saturday, October 24, 2026
cerebras.ai348 pts on HN
Cerebras Deploys Trillion-Parameter AI Models