Block Dessa Explained: Speech-Only Training, Local Hardware and Availability

Block introduced Dessa on October 23, a conversational AI model trained entirely on human speech. It can listen while speaking, rather than waiting for a clean handover between turns. For developers, this is an interesting research preview, with limited access rather than a public API launch.

Dessa lead Ragavan Thurairatnam’s announcement invites AI safety researchers to explore the model. That distinction matters if you arrived looking for a download or an integration guide.

What Dessa changes about voice AI

According to Block’s announcement, Dessa was trained from scratch without text or a pretrained model. The company says it handles conversational timing, interruptions and short acknowledgements, and runs locally on a MacBook Pro with an M4 Max or better.

Full-duplex means the system can receive and generate speech at the same time. The important research question is how much useful conversational behaviour can emerge directly from audio, without inheriting language knowledge from a text model.

Full-duplex voice itself is not new. Kyutai’s 2024 Moshi paper describes a system that models the user’s speech and its own speech in parallel. Moshi starts from a text-language-model backbone and predicts text tokens alongside audio. Dessa’s speech-only training claim is the more specific distinction.

That is worth keeping separate from a claim of better intelligence. Learning to take a turn naturally and answering a difficult question correctly are different capabilities. A voice model needs to be evaluated on both.

Dessa availability, API and pricing

At the time of writing, the public page offers recorded demonstrations and an early-access request for a small group of safety researchers. It says launch links are coming soon. The announcement provides no public API pricing, downloadable weights or general-release date.

Treat the M4 Max statement as a hardware claim, not an installation path. It does not establish which memory configurations work, how much processing headroom an application needs, or what simultaneous workloads it can support.

For a team choosing a voice stack today, those details decide whether a promising demo can become something dependable. Local execution could reduce dependence on a remote inference service, but it also moves capacity planning onto the device.

What developers should watch next

The useful test would involve conversations that are harder to stage: two people speaking over each other, a correction halfway through an answer, or a long pause that should not end a turn. Measure how often the model interrupts incorrectly, not just how quickly it starts talking.

Then check whether it preserves the corrected information. An assistant that sounds attentive but continues with the wrong instruction has improved the presentation without fixing the task.

Dessa makes speech-only training worth watching. Its next meaningful milestone will be evidence that natural conversation remains reliable outside the team’s selected recordings.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn launches and product updates into clear stories that developers and founders want to read and share.

Book a 30-minute call.