How we measure reduced-precision variants

A -fp4 model is the same model served at reduced precision (FP4), as a separate model id. Before one is published, Anoman runs one measurement: 100 synthetic prompts (25 JSON-shaped, 25 short classification, 25 Bahasa Indonesia questions, 25 short code tasks; never customer data), streamed through both routes in the same run, each with its live routing settings. The page shows the median output tokens per second of each route and the date. Results vary by provider and time; the numbers are not a guarantee. When the serving provider changes, the variant is flagged for re-measurement; until a new run, the date shown is the last one.