The Block Dessa AI model is a research preview, but its pitch is bold: it listens and talks at the same time, and Block says it learned from human conversation alone. Block's in-house AI lab unveiled the model on Friday, October 23, 2026. There is a waitlist, but no public release, API, app or price.
Most of the claims cannot be checked yet. Unite.ai describes Dessa as a true speech-to-speech model, meaning audio goes in and audio comes out. Many voice systems chain separate steps instead: speech becomes text, a text-based language model writes a reply, and that reply is turned back into speech.
The announcement also travelled fast. Jack Dorsey promoted it on X, calling it "the most authentic full-duplex speech model ever built." CellCog reports that his post drew about 4.1 million views. Reach says nothing about whether the model works, so this article separates what Block has said from what anyone has verified. Block is best known to consumers for Square and Cash App, and Dessa is a research effort separate from both.
What the Block Dessa AI model claims to do
Block calls Dessa full-duplex, which means it can listen while it speaks. Human conversation works that way. People interrupt, murmur "mm-hm" and start replies before the other person finishes. According to Unite.ai, Block says Dessa learned when to interject and when to give short acknowledgments from real conversation.
That addresses a familiar complaint. Most assistants wait for silence, respond, and stop talking only when they hear you speak over them. A model that listens and speaks at once could, in principle, cut those awkward gaps. Block has not published latency figures, so no one outside the company can say how much faster Dessa is.
Block frames Dessa as a new category rather than a faster version of an existing model, according to Tech Insider. The company also hedges its boldest line. As Unite.ai reports, "as far as the company knows, Dessa is the world's first AI that can hold a realistic live conversation after being trained only from human speech." That qualifier is careful. It means the "first" claim is Block's own judgment, not an independent finding.
How Block says it trained Dessa
According to Digg, Block says it trained the model from scratch on human conversation, without text or pretrained models. That is the core of the pitch. Most voice AI systems are built on top of text-based language models, and daily.dev describes Dessa as an exception to that pattern.
The hardware claim rests on thinner ground. CellCog says Block trained Dessa on a single NVIDIA GB200 NVL72 rack. No other source in this research confirms that figure, so it should be checked against Block's own page before anyone repeats it. A single rack would matter to anyone estimating what a model like this costs to build.
Dessa also raises a question about where training material comes from. The coverage reviewed for this story does not say how the conversations were collected, licensed or consented to. For a parallel debate over training sources, see OpenAI's decision to stop training frontier models on the live internet.
Conversation data carries real exposure risk. Meta paused Muse sharing after 6,800 shared chat links were exposed, a reminder that private exchanges can leak once they leave the system that created them. Any lab working with human conversation has to answer that problem before it asks the public to trust the data.
Why the demos and hardware claims need scrutiny
Block says the clips are live recordings. Commentators still question the selection, and that concern is fair, since Block chose which clips to publish. Digg quotes one reviewer who found some chats "a little awkward and uncanny." Community reaction has split between fascination with a voice-only, full-duplex model and heavy skepticism about whether it can handle facts, tool use or noisy real-world conversations.
Those are the tests that matter most for a voice product, and none appear in the published material. No benchmark, latency figure or model size has been released. Every performance claim therefore comes from Block itself. The table below sorts the claims by how they have been checked so far.
| Claim | Who reported it | Independent check |
|---|---|---|
| Speech in, speech out, full-duplex | Unite.ai | None found |
| Trained from scratch on conversation, no text | Digg | None found |
| Runs locally on MacBook Pro M4 Max or better | Unite.ai, Digg | None found |
| Single NVIDIA GB200 NVL72 rack for training | CellCog only | Not confirmed |
| Latency, model size, benchmarks | None published | Not available |
The MacBook requirement deserves a closer look. Block says Dessa runs locally on a MacBook Pro M4 Max or better. That points to a high-end laptop, not a phone or a budget machine, so the on-device promise reaches a narrow slice of users, at least on the stated hardware.
Who gains and who loses if the claims hold
Developers and voice product teams stand to gain the most. If on-device, full-duplex speech works as described, it could reduce the lag and awkward pauses that make current assistants feel slow. Teams building customer service bots, in-car voice controls or accessibility tools will watch closely. None can build on Dessa yet, because there is no API.
Consumers gain nothing for now. Dessa has no app, no download and no price, so the only action available is watching for waitlist access. Investors and fintech watchers should read the announcement as Block research rather than a product launch. It does not affect Square or Cash App revenue, and it gives no release date.
Competitors that build voice AI on text-based language models have the most reason to pay attention. If a model trained only on speech can hold a realistic conversation, the assumption that voice AI must start with a language model gets tested. According to Digg's account of the project page, Ragavan Thurairatnam, who leads Dessa's work, called the release "a GPT3 moment for speech."
That comparison is ambitious. GPT-3 became a reference point largely because developers could use it and judge the results for themselves. Dessa's waitlist means that test has not happened yet.
What to do now
No action is required, because Dessa is not available to use. Readers who want to evaluate voice AI can do that with tools they already have. A few steps help:
- Read Block's own announcement before relying on this coverage. Digg, daily.dev, CellCog and Tech Insider are secondary sources, and at least one figure is unconfirmed.
- Ignore download pages that claim to offer Dessa. The announcement lists no app or installer.
- If you want waitlist access, look for it on Block's own site rather than through third-party pages.
- Test the voice assistant you already use on the things critics flagged: interruptions, factual questions and background noise. That gives you a baseline for comparison once Dessa results appear.
- Developers should wait for published latency, model size and benchmark numbers before planning around the model.
What to watch next
Three signals matter most. The first is whether Block publishes latency, model size and benchmarks, which it has not yet done. The second is whether the waitlist expands beyond the small group of AI safety researchers Block has invited. The third is whether outside researchers test Dessa on factual accuracy, tool use and noisy audio.
The most useful evidence will probably come from people who did not build the model. The GB200 training figure should also be confirmed against Block's own materials. If Block backs it up, the story becomes one about cost and scale. If not, the company will have to make its case on demos alone.
