Block's Dessa: a full-duplex conversation model trained only on speech, with no text, released as a closed research preview

Block's Dessa is a voice AI model that the payments company says learned to talk from audio alone. Block calls it "a research preview of a full-duplex speech model," trained "from scratch on nothing but human conversation. Sounds! No text, no pretrained models." A Hacker News thread about the page started on October 23, 2026. It's described as small enough to run on a high-end laptop, which places it among small models, but nobody outside a small group of safety researchers can use it yet. That sets it apart from the voice launches in our GPT-Live-1.1 and Grok Voice 3.0 story, which shipped to paying users.

This story is based on Block's Dessa page, checked October 25, 2026. Block hasn't published a paper, model card or code.

The short version

• What it is: a speech-to-speech model, "Audio in, audio out," that is full-duplex: "it listens and talks at the same time."
• The claim: trained from scratch on recordings of human conversation, with no text and no pretrained models.
• Compute and hardware: trained on "only one NVIDIA GB200 NVL72"; runs locally on "a MacBook Pro M4 Max or better."
• Availability: none. Early access goes to "a small group of AI safety researchers." No weights, license, API or app.
• Not disclosed: parameter count, training hours, languages, benchmarks or how latency was measured.

What does "trained only on speech" mean?

Block says Dessa never saw text. Most voice assistants either chain separate steps (speech-to-text, a text language model, then text-to-speech) or, in newer speech-to-speech systems, start from a language model pretrained on text. Block says Dessa skips both: "No text, no pretrained models."

The page argues this is harder than it sounds, because "a single spoken word takes thousands of times more bytes than the same word in text." It also claims Dessa "learned language from hundreds of times fewer words" than text language models, without giving either figure.

What the page doesn't say is how audio is turned into something the model can learn from. It names no audio tokenizer or codec, and no amount of training audio. It also doesn't say whether transcripts were used anywhere, for example in filtering data or in evaluation. "No text" is Block's description, and without a paper it can't be checked.

How fast is Dessa, and what can it do?

Block says "Inference runs faster than real time, so there's no awkward wait." That is a vendor claim with no measurement attached: the page gives no response time in milliseconds, hardware for the measurement or test conditions.

The evidence offered is six recorded demos, labeled as live conversations with subtitles added afterward. They cover talking about the release, remembering a movie, an anniversary surprise, an expressive conversation, life advice and a conversation with OpenAI's GPT-Live-1. Block notes that the developer of GPT-Live-1 "is not affiliated with Block and has not reviewed or endorsed Dessa." Demos chosen by the developer show what a model can do on a good take, not how often it does it.

Why did Block build it?

Block gives three reasons: curiosity, better understanding between people and AI, and safety. On safety, the page says voice AI "is still nearly deaf to the emotion in our voices," and asks readers to imagine "being in serious pain while a voice AI has no idea and just keeps going." Block presents a model that learned from sound as a step toward hearing that kind of distress. It hasn't published any evaluation showing Dessa does.

It also frames the release as a reason for caution: "A model that learned to talk from sound alone is new territory, and we want outside eyes on it."

Who built Dessa?

The page credits a small AI research lab at Block, led by Ragavan Thurairatnam and reporting to Jack Dorsey, founded by former members of Dessa, a Toronto deep-learning startup. Square, now Block, bought that company in February 2020, according to BetaKit; Thurairatnam, a co-founder, was then its head of machine learning.

The team has history with lifelike speech. In May 2019, Dessa's engineers, including Hashiam Kadhim and Rayhane Mama, who are both credited on the new model, published RealTalk, a text-to-speech system that imitated podcast host Joe Rogan's voice. At the time they said "we will not be releasing our research, model or datasets publicly," citing risks such as impersonation scams.

Is Dessa open source?

No, not as of October 25, 2026. The page links to no weights, repository or paper, and names no license. It shows a "Download" button that says "Download coming soon" and a "Join the waitlist" button that says "The waitlist opens soon," and it says "Launch links are coming soon." Its privacy notice, effective October 21, 2026, covers the website, not the model.

By contrast, Mistral's earlier Voxtral speech models are Apache 2.0 open weights, while its new Voxtral 3 models are API-only, as our Voxtral 3 story explains. If Block releases Dessa to run on Macs, the memory math in our guide to small models for laptops will apply, but Block hasn't said how much memory it needs beyond naming an M4 Max.

What's confirmed, reported and uncertain

• Confirmed (Block's page): the "no text, no pretrained models" description, full-duplex operation, one GB200 NVL72 for training, the MacBook Pro M4 Max requirement, the research-preview status, the team and the safety-researcher early access.
• Vendor claims, not independently checked: "faster than real time," "hundreds of times fewer words," and the demos.
• Unknown: model size, training data volume and sources, languages, benchmarks, license, release date and whether weights will be published.