Tokenization and Token Costs

45 or 793 tok/s: the same model, and only one of those numbers sizes your box

A dev.to explainer on local inference benchmarks makes a point worth pinning up: tokens per second is a function of how many users you tested with, not a property of the hardware.