First-Token Limits

Which chip wins depends on how long a user may wait for the first token. For each cap on time to first token, this reads the measured configurations that also clear the interactivity floor and keeps the cheapest one per vendor. Measured rows only: no interpolation, and every bar links to the run behind it.

Benchmark Config

Model

Scenario

Precision

Chart Config

Cost Provider

Token Type

Minimum Median Interactivity (tok/s/user)

TTFT Caps (s, comma-separated)

First-Token Limits

First-token limits find, for each chip vendor, the cheapest measured config per million tokens under time-to-first-token (TTFT) caps such as 2 s, 5 s and 10 s at a chosen minimum interactivity. Every bar is a real run with no interpolation, traceable to its GitHub Actions run.