Token Speedometer

Pick a rate (or two) and watch what that generation speed actually feels like. Purely theoretical: no model, no endpoint, just a metered drip of pseudo-tokens (~4 characters each). Set a context size and the lanes sit silent through prefill first, exactly like a real endpoint.

Rule of thumb: 1 token is about 3/4 of a word; comfortable silent reading is ~250 words/min, roughly 5-6 tok/s. Time to first token = context / prefill rate (8192 tok at 305 tok/s prefill = ~27s of silence before the first word). Actual tok/s measures the decode phase only.