Evergreen Tools
|
← Back to AI Tools

Dessa Full-Duplex Voice Model

A speech model trained from scratch on human conversation that listens and talks at once, on your own laptop

Official Source

This page summarises an official announcement — check the vendor site for current availability and pricing.

Features

  • ✓ Full-duplex: listens and speaks at the same time, handling interruptions, laughter and tone shifts
  • ✓ Trained from scratch on speech alone, with no text and no pretrained language model underneath
  • ✓ Block says training fit in a single NVIDIA GB200 NVL72 rack
  • ✓ Block says it runs locally on a MacBook Pro with an M4 Max chip or better
  • ✓ Research preview only, with early access limited to a small group of AI safety researchers

How to Use

  1. Read Block's official announcement and listen to the demo clips
  2. Security researchers can apply for early access through the channel Block published
  3. If granted access, load the model locally on a MacBook Pro with M4 Max or better
  4. Test with real conversation, noting where the model misreads interruptions, laughter or emotion

FAQ

What is Dessa?

A full-duplex voice model announced on 23 October 2026 by the research lab at Block, the company behind Square and Cash App. Block says it was trained from scratch on human speech alone, listens and speaks at the same time, and runs locally on a MacBook Pro with an M4 Max chip or better.

How is Dessa different from a normal voice assistant?

Most assistants chain record, transcribe, LLM, then synthesise, so they wait for you to finish. Dessa is a speech-to-speech model, and Block stresses that it handles interruptions, laughter and tone changes, which makes the rhythm closer to a real conversation.

Can I download or call Dessa today?

No. Block keeps it as a research preview with early access for a small group of AI safety researchers. There is no public weight release, no API, and no published model size or benchmark results.

Why does Block emphasise local execution?

Block says the model runs on consumer hardware such as a MacBook Pro with M4 Max or better, so speech never has to leave the machine for cloud processing. Voice carries a lot of personal information, so local inference cuts the data exposure.

How much of what I read online about Dessa is verifiable?

Only what Block itself published, and only where it gave a number. Beyond a single NVIDIA GB200 NVL72 rack for training and running locally on M4 Max, Block disclosed no model size, benchmark score or third-party replication, so treat anything more specific as unverified.