Cerebras runs a 3-trillion-parameter model across 59 wafers at 2,000 tokens per second

Cerebras Deploys Trillion-Parameter AI Models

Cerebras explains how it deploys trillion-parameter models on its wafer-scale chips. Using the open-source Kimi K3 model with roughly 3 trillion parameters, it splits the model across 59 WSE-3T wafers, each holding 44 GB of on-chip memory. Because each layer fits entirely on a wafer, only activations move between chips via RDMA over Ethernet. The pipeline stays full with different requests, and KV cache size on dense layers caps concurrency.