Block’s Dessa previews local conversation without a text pipeline

Block is previewing Dessa, a conversational speech model designed to listen and speak simultaneously. Its public page describes training directly on human speech, without a text-based training stage or a pretrained model. These are Block’s claims; we have not run the model.

Public demonstrations: October 23, 2026, as reported by TechCrunch on October 24. This source review was published October 24.

What makes the preview interesting

The company describes full-duplex, audio-to-audio conversation that can respond to timing and expressive cues. It says local inference runs on a MacBook Pro M4 Max or better. That hardware statement is publisher guidance, not an AiBrainWorX benchmark or evidence of support for our RTX 3060 setup. Source: Block’s Dessa page.

Local capability does not establish download availability

The reviewed page offers a waitlist and limited early access for safety researchers. It says launch links are forthcoming. We found no public checkpoint download or open-weight license on that page, so this announcement does not yet establish an installable open-source option for our launcher.

Our take: test the conversation, not just the voice

For a future local voice workflow, we would measure interruption handling, turn timing, recovery from misunderstood speech and resource use alongside sound quality. A compelling demonstration helps identify what to test; it does not replace those measurements.

Test status: source review only. No model, audio or runtime was downloaded. Our current model library remains separate from this research preview.