Block's Dessa Is a Speech-to-Speech Model Trained Only on Human Speech. Early Access Is for AI Safety Researchers, and It Needs a MacBook Pro M4 Max.
Block published Dessa, a full-duplex speech-to-speech model it says was trained from scratch on human speech alone, with no text pretraining and no language-model backbone, running locally on a MacBook Pro M4 Max or better at faster than real time. Early access is limited to a small group of AI safety researchers, not a general preview. A demo has it conversing with OpenAI's GPT-Live-1 under an explicit non-endorsement disclaimer. Every capability claim is Block's; no independent evaluation exists. We read the Dessa page.
By T.W. Ghost
Block, the payments company, has a small AI research lab that reports to Jack Dorsey, and this week it published a speech model with an unusual pedigree. Dessa, at block.xyz/dessa, is a full-duplex speech-to-speech conversation model that Block says was trained from scratch on human speech and sounds only: no text, no pretrained language model underneath. It runs locally on a MacBook Pro with an M4 Max or better, faster than real time, and Block says it was trained on a single NVIDIA GB200 NVL72 system. The team includes former members of Dessa, the Toronto lab, which is where the name comes from.
What Block is offering is narrower than the launch coverage suggests. The page's headline is a call for safety researchers: limited early access for a small group of AI safety researchers, with a contact address, not a public research preview. There are no weights, no API, and no license for general use. The demo on the page has Dessa talking to OpenAI's GPT-Live-1, and Block adds that the developer of GPT-Live-1 "is not affiliated with Block and has not reviewed or endorsed Dessa." We read the page. That is the only primary source there is. Here is what held up.
What Block Says It Built
| Block Dessa | |
|---|---|
| What | Full-duplex speech-to-speech conversation model |
| Training (Block's claim) | Human speech and sounds only; no text; no pretrained model; one NVIDIA GB200 NVL72 |
| Runs on | MacBook Pro M4 Max or better, locally, faster than real time |
| Access | Early access for a small group of AI safety researchers; contact ragavan+dessa@block.xyz |
| Weights, API, license | Not offered |
| Independent evaluation | None published |
| Demo | Conversation with GPT-Live-1, with a non-affiliation and non-endorsement disclaimer |
The gate is the news. A model trained without text is a research object, and Block is treating it as one: outside safety eyes first, broader release unspecified. Do not expect to download it.
The hardware floor is high for a local model. M4 Max is the minimum Block names.
Everything about capability is a demo plus Block's prose. No benchmark, no third-party listen, no AA speech board entry. The GPT-Live-1 crossover is a demonstration, not a partnership, and Block's own disclaimer says so.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Full-duplex speech-to-speech; trained on human speech only; no text; no pretrained backbone | Vendor (Block's page) |
| Runs locally on M4 Max or better, faster than real time | Vendor |
| Trained on one GB200 NVL72 | Vendor |
| Early access limited to AI safety researchers | Verified wording (Block's page) |
| Weights, API, or license for general use | Not offered |
| GPT-Live-1 demo with non-endorsement disclaimer | Verified (Block's page) |
| Any independent evaluation | None |
Who Should Use It
Safety researchers: the email path is on the page. Everyone else: there is nothing to use yet, and no date for when there will be.
The frontier is elsewhere: on the speech-to-speech boards that exist, the leaders this month are Grok Voice 3.0, Gemini 3.8 Live, and OpenAI's GPT-Live-1. Dessa is not on any of them, by design. Our model comparison has the current standings side by side.
Sources
*Not sure whether a research-grade voice model is worth tracking before it is available? Take the free 2-minute quiz and get matched. Then read about the voice model Dessa talks to in its own demo: GPT-Live-1.1.*