October 25, 2026
Kimi K3 on Cerebras: claimed generation speed starts at 1,000 tokens per second

Cerebras split Kimi K3 across 59 WSE-3T wafers: 35 wafers for 69 linear-attention layers and 24 wafers for dense layers (Cerebras blog, Oct 22, 2026). Weights stay on the chips, only activations pass between wafers, and the RDMA hop takes about 2 microseconds.