“Cerebras Inference generated 969 output tokens/s when given a 1,000 token prompt.”
The 969 tokens a second headline comes from a 1,000 token prompt. Stretch the prompt to 100,000 tokens and it drops to 539. The record was set on “a customer workload” last week, and general availability waits until Q1 2025 at $12 per million output tokens. A speed record you can’t buy yet is a sales pitch.