Ackren answers in 3.6 milliseconds

The figure is wall-clock time from input received to the last word of an eighty-word reply, on a single core, with the artefact already mapped. It excludes network, because there is no network: the engine runs on the device.

The comparison figure is a hosted transformer answering the same prompt at the same length. The ratio is approximately 7,700 to one. Both numbers and the harness that produced them go out together, because a ratio without its method is not a measurement.

This is not a tuning result and it does not erode under load. There is no statistical inference in the pipeline, so there is no matrix multiplication whose cost grows with model size, and no batching regime whose latency degrades as the queue fills. The work is lookups and rule application, and the cost is bounded by the grammar rather than by the hardware budget.

Latency is the wedge, and it is the wedge because voice interfaces break past roughly 200 milliseconds and nobody solves that with money. A hosted model that is fast today is fast because someone is paying for headroom.