Cerebras Claims Kimi K3 Hits 2,000 Tokens per Second on 59 Wafers
Summary
Cerebras says it spreads Kimi K3, an open-source 3-trillion-parameter model, across 59 of its WSE-3T wafers, each holding 44 GB of on-chip memory. One dense layer sits on each of 24 wafers and two linear-attention layers on the other 35. Cerebras claims 1,000 to 2,000 or more tokens per second.
Key Points
- Cerebras reports Kimi K3 scores 88.3% on Terminal-Bench 2.1, just under GPT-5.6 Sol's 88.8%.
- Cerebras puts each inter-wafer hop at roughly 2 microseconds using RDMA over Ethernet.
- Concurrency is capped by the KV cache space left on each dense-layer wafer.