Block's speech-only model shifts to a local M4 Max path
A local M4 Max path is the change in Block's launch page for Dessa. It names a MacBook Pro M4 Max or better as the hardware floor.
| Dessa | |
|---|---|
| hardware | MacBook Pro M4 Max or better |
| inference | faster than real time |
| training source | nothing but human conversation |
| language learning | hundreds of times fewer words |
| recording format | two synchronized speaker tracks |
The launch page describes Dessa as a speech-to-speech model trained from human speech. It says the model can run locally on a MacBook Pro M4 Max or better.
The Verge describes the demos as awkward and uncanny. The vendor page points to live recordings and two synchronized speaker tracks as evidence that the exchanges are unsteered.
The useful change for a developer is that a speech-only model can sit on a laptop instead of a cloud endpoint. The constraint for a chip buyer is that M4 Max or better is the only named floor.
The page says inference runs faster than real time, but it does not give milliseconds, parameter count, dataset hours, or access terms.
Model size, training corpus, latency, price, and independent benchmark remain unmeasured.
The buyer is constrained by local hardware, not by the model's claimed personality.
After @RagavanThurair. We did not report this. The pictures, if any, are theirs.
