Most voice assistants don’t really listen to you. They turn your speech into text, think in text, and read an answer back. Block, the company behind Square and Cash App, says its new model skips the text entirely. Block Dessa is a speech-to-speech AI that the company says was trained from scratch on nothing but recorded human conversation: no text, no pretrained language model. Block unveiled it on Friday, October 23, on a launch page, with Dessa lead Ragavan Thurairatnam calling it “a GPT3 moment for speech” in his announcement on X at 12:38 PM ET.
The demos have spread fast. Jack Dorsey’s launch post had about 12.8 million views by Sunday morning, according to X’s counter. But you can’t use Dessa. It’s a research preview with no app, no API and no price, and the only people getting in early are a small group of AI safety researchers.
- What it is: a full-duplex speech-to-speech model (audio in, audio out, listening while it talks) that Block says learned language only from human speech.
- Price and availability: none yet. It’s a research preview with a waitlist. Block hasn’t published an app, an API, a price or model weights.
- Hardware: Block says it was trained on a single NVIDIA GB200 NVL72 rack and runs faster than real time on a MacBook Pro with an M4 Max chip or better.
- The catch: everything public so far is Block’s own demo recordings. There are no benchmark scores, no parameter count and no outside testing yet.
What Block Dessa is, and what “trained only on speech” means
Block’s page describes Dessa as “a true speech-to-speech model” that is full-duplex, meaning it can listen and talk at the same time instead of waiting for you to finish. Because it learned from real conversations, Block says, it knows when to cut in and when to drop in a “yeah!” or “got it!” to show it’s following. It “natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh,” the page says, the kind of signal that disappears when speech is converted to text first.
Full-duplex voice AI isn’t new on its own. Kyutai’s Moshi, released in 2024, also listens and speaks at once, but its research paper describes a model built on top of a text language model. Block’s claim is the training recipe: “No text, no pretrained models.” The company says that, as far as it knows, Dessa is “the world’s first AI you can have a realistic live conversation with that was trained only from human speech.” That’s Block’s claim, and no outside group has checked it yet.
Meet Dessa, our AI model trained from scratch on nothing but speech.
We believe speech is the natural interface for technology. So we built Dessa.
And we’re inviting AI safety researchers to explore Dessa.
Learn more: https://t.co/6wA8KhWdjn pic.twitter.com/vXM43uUFzx
— Ragavan Thurairatnam (@RagavanThurair) October 23, 2026
Video: Block. “Meet Dessa,” a three-and-a-half-minute launch video posted on X by Block’s Dessa lead (@RagavanThurair) on October 23, 2026 (watch on X). It mixes team members talking with Dessa and animated clips of its conversations.
How small is it? What Block says about the hardware
The size claims are the most surprising part. Block says it trained Dessa on “only one NVIDIA GB200 NVL72,” a single rack that Nvidia lists with 72 Blackwell GPUs. Frontier text models are typically trained on far larger clusters. Block also says inference “runs faster than real time” and that Dessa “runs locally on a MacBook Pro M4 Max or better.” Dorsey put it as “small enough to run on your macbook,” and Thurairatnam’s post says it runs “on consumer GPUs.”
On data, the page says Dessa “learned language from hundreds of times fewer words” than LLMs need. What Block hasn’t published is just as important: the parameter count, hours of training audio, which languages it handles, measured response latency, and any benchmark scores. Until those numbers appear, the efficiency story is a company claim.
What the Dessa demos show, and what they don’t
Block posted more than 20 recordings. It says every clip is “a live recording of a person talking to Dessa” and that “nobody steers it”: the model picks what to say, when and how, including its own voice and personality. Subtitles were added afterward. In one clip it laughs mid-sentence; in another, asked to speak up, it gets louder, which Block says “nobody taught it.” In a third, a person blanks on a movie title and Dessa finishes the sentence.
One demo pairs Dessa with GPT-Live-1, which Block’s page also calls ChatGPT Live. Block labels both voices AI-generated and says that model’s developer isn’t affiliated with Block and hasn’t reviewed or endorsed Dessa.
These are still clips Block chose to publish. There’s no independent leaderboard result yet, so there’s no apples-to-apples way to compare Dessa with voice models people can actually pay for today. And the demos show the limits of learning facts from sound: in one of the extra samples Dessa calls itself “an open source AI,” and in another it agrees when a person says so. Block hasn’t announced any plan to release Dessa’s weights.
Who built Dessa, and why Block calls it a safety project
The model comes from what Block calls “a (very) small AI research lab” founded by former members of Dessa, the Toronto deep learning startup the model is named after. The team says it has built generative speech models for over eight years and reports directly to Dorsey. The page lists six researchers and five current and former interns.
Block gives three reasons for the project: curiosity, better communication between people and AI, and safety. Its argument is that today’s AI is “nearly deaf to the emotion in our voices,” so a voice agent could keep going while someone on the line is in serious pain. That’s also why the first outside users are safety researchers. Block’s post on X invites “a small group to explore Dessa” and help it understand “what it can and can’t do.”
How to try Dessa right now
For most people, you can’t yet. The launch page has a “Join the waitlist” button, and safety researchers can apply through a form that asks for a name, an institutional email address and, optionally, a Google Scholar or blog link. The page also has a “Download” button, but in the page code we checked on Sunday it was wired to a “Download coming soon” message. Block hasn’t given a date, a price or a list of supported devices beyond the M4 Max MacBook Pro claim.
Frequently asked questions
What is Block Dessa?
Dessa is a full-duplex speech-to-speech AI model from Block, announced October 23, 2026. Block says it was trained from scratch only on recorded human conversation, with no text and no pretrained models, so it can hear and respond to tone, laughter and pauses.
Can I use Dessa, and is it free?
Not yet. Dessa is a research preview with a waitlist, and early access is going to a small group of AI safety researchers. Block hasn’t published an app, an API or any pricing.
Is Dessa open source?
No. Block hasn’t released Dessa’s weights or code or announced plans to. The page’s Download button was marked “coming soon” when we checked on October 25, 2026.
What hardware does Dessa need?
Block says Dessa runs faster than real time locally on a MacBook Pro with an M4 Max chip or better, and that it was trained on a single NVIDIA GB200 NVL72 rack. It hasn’t published the model’s size or other system requirements.
Sources
- Block, “Dessa – the Conversation Model trained only on speech” launch page (October 23, 2026; training claims, full-duplex and hardware claims, demo recordings and GPT-Live-1 disclaimer, team, safety-researcher form, waitlist and Download buttons): block.xyz/dessa
- Ragavan Thurairatnam (Dessa lead) on X, announcement (October 23, 2026, 12:38 PM ET; “GPT3 moment for speech,” “consumer GPUs”): x.com/RagavanThurair/status/2113671324400861645
- “Meet Dessa” launch video posted on X by Block’s Dessa lead, @RagavanThurair (October 23, 2026, 1:59 PM ET): x.com/RagavanThurair/status/2113691903866872251; Block (@blocks) on X, safety-researcher invitation (October 23, 2026, 3:55 PM ET): x.com/blocks/status/2113721026324615173
- Jack Dorsey (@jack) on X, launch post and reply (October 23, 2026, 2:23 PM ET; view count as shown October 25): x.com/jack/status/2113697902916751661
- Kyutai, “Moshi: a speech-text foundation model for real-time dialogue” (arXiv, 2024): arxiv.org/abs/2410.00037
- NVIDIA, GB200 NVL72 product page: nvidia.com/en-us/data-center/gb200-nvl72
All sources accessed on the morning of October 25, 2026 (ET).
Last updated: October 25, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
