Measured, not promised.
Published results from the standard benchmark suite, run on real sessions in this lab.
Loading published results
How the standard suite measures
A short answer for time to first token, a sustained 1,000-token answer for decode speed, synthetic documents of 4K, 16K and 64K tokens (optionally 256K and 1M) with a counting question for prompt processing, and eight deterministic spot-check tasks. Every request runs one at a time at temperature 0 through the same session path customers use. Power is sampled from the machines' telemetry during the run.