Cerebras Inference — models and pricing

Cerebras Inference runs open-weights models on the Wafer Scale Engine — a single-chip accelerator the size of a dinner plate — rather than clusters of GPUs. The architectural payoff is order-of-magnitude faster token generation: observed tokens-per-second on supported models routinely sits in the 1500–2500 range, with median time-to-first-token under 200ms.

The catalog is intentionally narrow: a curated set of Llama 3.x, Qwen 3, and Mistral models tuned for the WSE rather than an exhaustive open-weights menu. Pricing skews higher per token than commodity GPU hosts, reflecting the latency premium and limited supply.

Best deployed for real-time voice agents, low-latency chat UIs, and agentic loops that make many sequential small calls — workloads where end-to-end latency, not per-token cost, drives the bill.