Block Dessa: How Jack Dorsey's Lab Taught An AI To Talk Without Text, And Why GoldieBench Can't Score It

Block Dessa is a speech-to-speech AI model from Block, the company behind Square and Cash App, and it learned to hold a conversation only by listening to people talk.

A very small research lab at Block built it, and that lab reports directly to Jack Dorsey, Block's cofounder and chairman.

Block's official Dessa page says it runs locally on a MacBook Pro M4 Max or better.

On 24 October 2026 that page showed Download and Join the waitlist buttons, with the launch links marked as coming soon and no price listed.

I'm Julian Goldie, I run GoldieBench, and I haven't used Dessa myself, so this post sticks to what Block and its team have published.

I'll cover who built it, how it learned to talk, what you can use today and what it means for voice AI, 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 pairing Hermes Agent with ElevenLabs, which is the separate voice layer Dessa is designed to do without.

Block Dessa at a glance

This table is the whole story on one screen.

What did Block announce?

Block announced Dessa as a conversation model trained only on speech.

It's a true speech-to-speech model, which means everything it knows came from human speech rather than written text.

It's full-duplex, so it listens and talks at the same time and knows when to cut in or make small sounds that show it's following you.

Block says it natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh and a voice cracking with sadness.

The page also says inference runs faster than real time, so there's no awkward wait before it answers.

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.

Ragavan Thurairatnam described it on X as a "full-duplex speech-to-speech model that you can run on consumer GPUs".

Jack Dorsey made a similar point on X when he said it's small enough to run on a MacBook.

Who is behind Block Dessa?

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

That's where the model gets its name.

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

It says it chases frontier problems that others ignore, and it reports directly to Jack Dorsey.

The page lists six researchers, three interns and two former interns, all in alphabetical order.

So the right way to read this news is as a long-standing speech research team that now sits inside Block, rather than a payments team trying AI for the first time.

How Block Dessa learned to talk without text

This is the part of the story that makes Dessa different from almost every AI model you've used.

Most AI models learn language from huge amounts of written text, and voice gets added afterwards.

Block's lab started from the other end, with the idea that babies learn to talk without ever seeing a word written down.

The team trained Dessa from scratch on nothing but recorded human conversation.

Block also says Dessa had to learn far more than language, including accents, pauses, breathing, sarcasm, humour, inflection and when to cut in.

The page admits that doing this without text to lean on is brutal, and then says the team figured it out.

Block also says Dessa knows a surprising number of facts, considering every fact came from raw sound.

What you can use today

When I checked the page on 24 October 2026, only part of it was open.

You can listen to six headline demo clips and 15 more samples, and Block says each one is a live recording that nobody steers.

The headline clips show Dessa laughing mid-sentence, getting louder when asked, finishing a forgotten film title, gasping at a surprise, showing a wide range of tone and sounding empathetic.

One clip shows Dessa talking with GPT-Live-1, which Block identifies as ChatGPT Live, and Block notes that its developer hasn't reviewed or endorsed Dessa.

The Download and Join the waitlist buttons are on the page, but the launch links were marked as coming soon.

A small group of AI safety researchers can request early access through a form that asks for an institutional email.

Block says more recordings will be shared on X soon.

Why Block says it built Dessa

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

It argues that today's voice AI flattens what we say into text and can't keep up with human timing.

It also says AI keeps gaining power and independence while staying nearly deaf to the emotion in our voices.

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

The page doesn't say Dessa will go into Square, Cash App or any other Block product, so I won't claim it will.

What Block Dessa means for voice AI

Most voice agents today, including my own, are a chain of speech to text, a text model and text to speech.

My Apollo copilot in the Agent OS works that way, with a wake word, a brain that answers or hands jobs to an agent, and a butler voice that replies.

Those agents are great at doing jobs, such as opening apps and building pages, but the text step loses the tone in your voice.

Dessa attacks exactly that weakness, because it never turns your voice into text in the first place.

Block's page doesn't mention tools or actions, so Dessa on its own wouldn't replace an agent that does the work.

The likely future is a natural voice like Dessa in front, with an agent doing the jobs behind it.

Block hasn't published an API or developer docs, so nobody can build that combination yet.

The limits of the Block Dessa launch

If you join the waitlist, the privacy notice says Block may share information with group companies such as Square and Cash App, and that it doesn't sell your information.

What to watch next from Block

These are the signals I'd track over the next few weeks.

The first is the launch links, because the Download and Join the waitlist buttons are the only routes to the model the page shows.

The second is Block's posts on X, because the page promises more recordings there soon.

The third is a price and licence, because nothing about cost or usage rights is published yet.

The fourth is an API or SDK, because that decides whether developers can connect Dessa to agents.

The fifth is any report from the safety researchers Block invited in, because they're the first outside people with access.

When those arrive, I'll know whether Dessa is a research milestone or a tool my members can actually use.

Why Block Dessa isn't on GoldieBench

Block 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.

The methodology page explains exactly how that scoring works.

Every model on the board reads a written prompt and writes code back, which is a text job from start to finish.

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

That isn't a mark against Dessa, because it's built for spoken conversation rather than building software.

It does mean you should never compare a speech model and a coding model on the same leaderboard.

The text models a voice agent still needs

If you want a voice agent that builds things today, the voice layer hands the job to a text model, and that's where GoldieBench helps.

These are real averages from the live board, and they measure one-shot builds, not voice quality.

Dessa would be the voice and one of these would be the builder, if Block ever opens a way to connect them.

The local models board shows which models run on your own hardware, and my best Hermes Agent model post explains how I pick a main brain.

For the product walkthrough and every demo clip, read my post on Dessa AI.

Also On Our Network

• Agent Operatorsseven things operators should know about Block Dessa
• AI Income Deskwhy a payments giant built Block Dessa
• AI Tool VerdictBlock Dessa against today's voice AI, with my verdict
• agentos.guideBlock Dessa explained for Agent OS users
• agentos.guideHermes Apollo, the voice engine I run 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 a second pair of eyes on your whole agent and voice setup, my team can review it in a free SEO strategy session too.

Use the voice agents you have today, listen to the demos and watch for the launch links, because that's the smart way to follow Block Dessa.