Cerebras Deploys Trillion-Parameter AI Models

Cerebras just broke the trillion-parameter barrier, running a 3-trillion-parameter Kimi K3 model across 59 wafers. Each 44-GB wafer holds about 16-GB of a layer, leaving room for a KV cache on dense layers. With 24 dense and 69 linear layers, the system stitches 59 wafers into a pipeline that streams tokens in microseconds. A 256k-token request, 4-bit quantized, fits in ~16 GB, and an 8-bit KV cache keeps latency low. The pipeline runs hundreds of requests, limited only by the 44-GB SRAM that holds KV caches on dense wafers. Each token hops between wafers in about 2 µs, giving a throughput of 1,000-2,000 tokens per second per request. This approach lets Cerebras outpace GPT-5.6 Sol on some benchmarks, with Kimi K3 scoring 88.3% on Terminal-Bench 2.1. Cerebras's 44-GB wafers are the new gold standard for AI.