Dessa AI Explained: What Block's Speech-Only Model Does, How To Get It, And Why GoldieBench Can't Score It

Dessa AI is a full-duplex speech-to-speech model from Block, and it learned to talk only by listening to recordings of human conversation.

Block, the company behind Square and Cash App, built it in a very small research lab that reports directly to Jack Dorsey.

Block's official Dessa page sums up the training in five words, which are "No text, no pretrained models".

The page says it runs locally on a MacBook Pro M4 Max or better, and on 24 October 2026 its Download and Join the waitlist links were marked as coming soon.

I'm Julian Goldie, I run GoldieBench, and I haven't used Dessa myself yet, so everything here comes from Block's own page.

I'll explain what it is, how it works, how to get access and what it can do, and then I'll explain why it isn't on the GoldieBench board.

I don't have a Dessa video, so the first video on this page is my one on Hermes voice agents, which use the older speech-to-text approach Dessa is trying to replace.

Dessa AI at a glance

This table is the whole model on one screen, taken only from Block's page.

What is Dessa AI?

Dessa AI is what Block calls a conversation model trained only on speech.

Most AI models learn from enormous amounts of written text, and voice is bolted on afterwards.

Dessa learned the other way round, by listening to people talk, which is how babies learn before they can read.

Block's claim is that, as far as it knows, Dessa is the first AI you can have a realistic live conversation with that was trained only from human speech.

It's full-duplex, which means it listens and speaks at the same time instead of taking turns like a walkie-talkie.

Block says it knows when to cut in and when to make the small backchannel sounds people use to show they're listening.

It also hears and speaks emotion natively, including sarcasm, laughter, a frustrated sigh and a voice cracking with sadness.

How Dessa AI works differently

Most voice AI today is a chain of three parts, which are speech to text, a text model that writes a reply, and text to speech.

Block's argument is that the first step throws away information that text alone can't carry, such as how you feel.

Block uses a clever example, which is saying the same sentence several times while stressing a different word each time.

The words never change, but the meaning does, and only a model that hears the sound can tell the difference.

How to get access to Dessa AI

There are three routes on Block's page, and only the first is fully open today.

First, you can listen to every demo recording on the page now, and Block says each one is a live conversation that nobody steers.

Second, you can use the Download and Join the waitlist buttons once the launch links go live, which the page said were coming soon on 24 October 2026.

Third, AI safety researchers can request early access through a form that asks for a name, an institutional email and an optional Google Scholar or blog link.

The page doesn't publish a price, a plan, an API or a release date for the download.

Every behaviour in the Dessa AI demos

Block says Dessa picks what to say, when to say it and how, down to its own voice and personality, and the subtitles were added afterwards.

The page lists 15 more samples underneath, with names such as Refund, Safety lab, Marvel gaming and Wise Dessa.

Block says nobody taught Dessa to get louder when asked, and that's the behaviour I'd test first.

The numbers behind Dessa AI

Block says Dessa learned language from hundreds of times fewer words than large language models need.

The page also says one spoken word takes thousands of times more bytes than the same word written down.

Speech never repeats exactly, so the same sentence said twice still produces different data.

Block says it trained the model on only one NVIDIA GB200 NVL72 despite all of that.

The page doesn't publish the model's size or its memory needs, so I'm not going to guess them.

Why Block built Dessa AI

Block gives three reasons, which are curiosity, better understanding between humans and AI, and safety.

The page says AI keeps gaining power while staying nearly deaf to the emotion in our voices.

Its example is a person in serious pain while a voice AI has no idea and just keeps going.

The lab is giving a small group of AI safety researchers early access to find risks it hasn't anticipated.

The limits of Dessa AI today

None of these limits are reasons to ignore Dessa, but they are reasons to wait before you build anything on it.

I'd rather tell you what the page leaves out than fill the gaps with guesses.

Who is behind Dessa AI

Block describes the team as a very small AI research lab, and the founders are former members of Dessa, the Toronto deep learning startup.

The lab says it has built deep learning models since 2012 and generative speech models for over eight years.

It reports directly to Jack Dorsey, whom the page lists as Block's cofounder and chairman.

Ragavan Thurairatnam signs the page as Dessa lead, and the page names six researchers plus current and former interns.

Block also gives the lab's own description of its focus, which is chasing frontier problems that others ignore.

What I'd test first when Dessa AI opens

I haven't been able to run Dessa yet, so this is my checklist for the day the download goes live.

I'd interrupt it mid-sentence to see how it handles being cut off.

I'd ask it to speak up and then speak quietly, because that's one of Block's headline behaviours.

I'd tell it a joke with a weak punchline to see whether it laughs or groans.

I'd forget a word on purpose to see whether it fills the gap like it does in the movie demo.

I'd sound frustrated on purpose to see whether its tone changes.

I'd time how quickly it replies on the MacBook Pro the page names, because faster than real time is Block's claim and not yet mine.

Is Dessa AI on GoldieBench?

No, Dessa isn't on GoldieBench, and it couldn't be scored there in its current form.

GoldieBench gives every model the same one-shot build tasks, such as browser games, simulations and visual scenes, and then judges the rendered result.

You can read exactly how that works on the GoldieBench methodology page.

Every model on the board is a text model that reads a written prompt and writes code back.

Dessa takes audio in and gives audio out, and Block says it was trained with no text at all, so it can't take a written build prompt or return a code file.

That doesn't make it better or worse than the models on the board, because it's built for a completely different job.

GoldieBench measures how well a model builds things, and Dessa is built to hold a natural spoken conversation.

The text models that would sit behind a voice agent

If you want a voice agent that actually does work today, the voice layer still hands the job to a text model, and that's where GoldieBench numbers help.

These are real averages from the live board, and they measure one-shot builds rather than voice quality.

A speech model like Dessa could one day handle the talking while a model like these handles the building.

Block hasn't published a way to connect Dessa to other systems, so that's a future setup rather than one you can build today.

The local models board shows which models run on your own hardware, and my post on the best Hermes Agent model goes deeper on choosing a main brain.

For the company and news side of this launch, read my post on Block Dessa.

Also On Our Network

• Agent Operatorshow to try Dessa AI and what to test first
• AI Income Desksix business uses for Dessa AI worth watching
• AI Tool Verdictevery Dessa AI feature rated on the evidence
• agentos.guidethe Dessa AI quick start for an Agent OS
• agentos.guideApollo, the voice copilot I use in my Agent OS today

If you run an SEO or agency business and want to talk through where voice AI fits, you can book a free SEO strategy session with my team.

If you'd like someone to review your whole agent and voice stack, my team can do that in a free SEO strategy session as well.

Listen to the demos, check your Mac and come back for the launch links, because that's the smartest way to get ready for Dessa AI.