Dessa AI is a full-duplex speech-to-speech model from Block, and it's built differently from almost every voice AI you've used because it never touches text.
A very small research lab at Block, the company behind Square and Cash App, trained it from scratch on human conversation audio with no pretrained models.
Block's page says it runs locally on a MacBook Pro M4 Max or better, with inference faster than real time.
On 24 October 2026 the page showed Download and Join the waitlist buttons and said launch links are coming soon, with no price or API listed.
I haven't tested Dessa myself, so this is a feature-by-feature reference built only from Block's official page and its team's posts on X.
Every rating below is an early score on published evidence, and I'll rescore it once I can run it.
I don't have a Dessa video, so here's my video on Hermes Jarvis, the voice agent I compare it against later in this post.
The Dessa AI Verdict In One Table
These are my early ratings, scored only on what Block's page and its team have published.
| Feature | Early score | Why it gets that score |
|---|---|---|
| Conversation feel | 9/10 | The demos show laughing mid-sentence, interrupting and finishing a forgotten film title. |
| Emotion handling | 9/10 | Block says it natively hears and speaks emotion, sarcasm, laughter and sighs. |
| Turn-taking | 9/10 | It's full-duplex, so it listens and talks at the same time and knows when to cut in. |
| Speed | 8/10 | Block says inference runs faster than real time, but I haven't timed it myself. |
| Local running | 8/10 | The page says it runs on a MacBook Pro M4 Max or better, which is high-end but still a laptop. |
| Training efficiency | 8/10 | Block says it trained on one NVIDIA GB200 NVL72 and needed far fewer words than a language model. |
| Safety approach | 7/10 | Block is inviting outside safety researchers, which is a good sign at this early stage. |
| Developer access | 2/10 | The page lists no API, no SDK and no integration docs. |
| Availability and price | 2/10 | The download and waitlist links were marked as coming soon, and no price is listed. |
My overall early verdict is that Dessa is the most interesting voice research launch I've seen this year, but it isn't a product you can build on yet.
🔥 Want my exact Is Dessa AI Any Good setup, step by step? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom
What Dessa AI Is
Dessa AI is a conversation model trained only on speech, which Block describes as audio in and audio out.
Everything it knows came from human speech, so Block calls it a true speech-to-speech model.
Block's claim is that, as far as it knows, Dessa is the first AI you can have a realistic live conversation with that learned only from human speech.
Ragavan Thurairatnam, who signs the page as Dessa lead, went further on X and called it a "GPT3 moment for speech".
That's a big claim, and it's his claim, so I'd wait for outside testing before repeating it as fact.
Feature 1: No Text, No Pretrained Models
Most voice AI uses a language model that was trained on huge amounts of written text.
Dessa was trained from scratch on human conversation audio and nothing else.
Block says Dessa learned language from hundreds of times fewer words than large language models need.
The lab is fair about the comparison, and it points out that language models also have to learn programming languages that Dessa doesn't.
Block also says Dessa knows a surprising amount of facts for a model whose every fact came from raw sound.
The page doesn't put a number on how much it knows, so I can't score its knowledge yet.
Feature 2: Full-Duplex Conversation
Full-duplex means the model listens and talks at the same time, the way you do on a phone call.
Block says Dessa learned from real conversations, so it knows when to cut in and when to backchannel with small sounds that show it's following you.
Most voice bots take turns like a walkie-talkie, and that's why they feel slow and awkward.
This is the feature that makes the demos sound so human.
Feature 3: Native Emotion
Block says Dessa natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh and a voice cracking with sadness.
The page argues that running speech through text strips all of that away.
Block uses a neat example, which is saying the same sentence several times while stressing a different word each time.
The words stay the same, but the meaning changes every time, and only a model that hears the audio can tell the difference.
Feature 4: Speed And Local Running
Block says inference runs faster than real time, so there's no awkward wait before it answers.
The page also says Dessa runs locally on a MacBook Pro M4 Max or better.
Jack Dorsey made the same point on X when he said it's small enough to run on a MacBook.
The page doesn't publish the model size, memory needs or supported operating systems, so I've held the score at 8 instead of 10.
Feature 5: The Training Run
Block says it figured out how to train Dessa on only one NVIDIA GB200 NVL72.
That matters because speech is far heavier data than text.
The page says a single spoken word takes thousands of times more bytes than the same word in text.
It also says speech never repeats, because even the same sentence said twice comes out as different data.
🔥 Want a voice agent you can actually use today? Inside the AI Profit Boardroom, I've got the Agent OS with my Hermes voice setup and step-by-step tutorials on wiring voice into real workflows. You'll be learning alongside 3,400+ members with four coaching calls every week. → Get access here
Feature 6: The Demo Recordings
The demos are the only hands-on evidence anyone has, so I've reviewed what each one is meant to show.
Block says every clip is a live recording, nobody steers Dessa, and the subtitles were added afterwards.
| Demo | What it's meant to show | My take |
|---|---|---|
| Talking about the release | It laughs mid-sentence and gets louder when asked to speak up. | Block says nobody taught it to get louder, which is the most striking claim on the page. |
| Remembering a movie | It finishes a film title the person has forgotten. | This shows knowledge and timing working together. |
| An anniversary surprise | It gasps under its breath and answers with what sounds like teary gratitude. | This is the clip that best shows emotion in its voice. |
| Expressive conversation | It shows a huge range in intonation and speed. | This shows it isn't stuck in one flat tone. |
| Life advice | Block says it sounds genuinely empathetic. | This is the clip coaches and support teams should hear. |
| Dessa in conversation with GPT-Live-1 | It talks to ChatGPT Live, and both voices are AI-generated. | Block notes the developer of GPT-Live-1 isn't affiliated with Block and hasn't endorsed Dessa. |
The page lists 15 more samples, including Refund, Safety lab, Zelda discussions and Wise Dessa.
Block says more recordings will be shared on X soon.
Feature 7: Safety Research Access
Block gives three reasons for building Dessa, which are curiosity, better understanding between humans and AI, and safety.
The page says AI keeps gaining power and independence while staying nearly deaf to the emotion in our voices.
Block is giving a small group of AI safety researchers early access to find risks the team hasn't anticipated.
The request form asks for a name, an institutional email and an optional Google Scholar or blog link.
I score this a 7 because inviting outside eyes is good practice, but the page doesn't yet publish any safety findings.
Feature 8: Developer Access
This is where Dessa scores badly today.
The page lists no API, no SDK, no documentation and no integration guide.
It also doesn't mention tools, function calling or any way to let Dessa take actions.
For anyone building agents, that's the feature that decides whether a model is useful, so I've scored it 2 until Block publishes something.
Feature 9: Availability And Price
The page shows Download and Join the waitlist buttons, and on 24 October 2026 it said launch links are coming soon.
There's no price, no plan list and no release date for the download.
That's why it gets a 2, and the score will jump the day Block opens the doors.
Dessa AI vs A Typical Voice Agent
Here's how Dessa compares with the voice agents I already run, such as Hermes Jarvis in my Agent OS.
| Area | Dessa AI | My Hermes voice agents |
|---|---|---|
| How speech is handled | Audio goes in and audio comes out of one model. | Speech is turned into text, a language model replies and a voice reads it back. |
| Tone and emotion | Block says it hears and speaks emotion natively. | They hear the words, but tone is mostly lost in the text step. |
| Doing tasks | The page doesn't mention tools or actions. | They open apps, build pages and remember things. |
| Where it runs | It runs on a MacBook Pro M4 Max or better. | They run inside my Agent OS on my own Mac. |
| Can I use it today? | The download links were marked as coming soon. | I use them every day. |
The honest summary is that Dessa wins on how the conversation feels, and my agents win on getting things done.
If you want to see what my setup looks like, read my post on the free Hermes Agent OS.
If you're choosing a text model to power an agent instead, my list of the best AI models for coding is a better fit, because Dessa isn't a text model.
If you want a second opinion on which AI tools fit your business, you can book a free SEO strategy session with my team.
Who Dessa AI Is For Right Now
Right now, Dessa is for three groups of people.
- AI safety researchers can apply for early access through the form on Block's page.
- Voice AI builders should study the demos, because they show where the bar is moving.
- Business owners should listen and plan, but not build anything client-facing until there's a price and an API.
For the news story, the team and why a payments company built a speech model, read my post on Block Dessa.
🔥 Want my honest tool verdicts and the setups behind them? The AI Profit Boardroom has the Agent OS, AI tool training and four weekly coaching calls where you can ask me which tools are worth your time. → Join the AI Profit Boardroom
Related Reading
- Block Dessa, the company story and what it means for voice AI
- The free Hermes Agent OS and what's inside it
- The best AI models for coding, ranked
Also On Our Network
- 🌐 The operator walkthrough for trying Dessa AI
- 🌐 Six business uses for Dessa AI worth watching
- 🌐 A Dessa AI quick start and how it fits an Agent OS
- 🌐 Dessa AI explained, and why it isn't on the GoldieBench board
FAQ: Dessa AI Review
Is Dessa AI any good?
On Block's published demos it sounds impressive, but nobody outside Block has tested it publicly, so I rate it one to watch rather than one to buy.
What makes Dessa AI different from other voice AI?
It was trained only on human conversation audio, it's full-duplex, and Block says it hears emotion that text pipelines lose.
Does Dessa AI run locally?
Block's page says it runs locally on a MacBook Pro M4 Max or better.
Can Dessa AI use tools like an agent?
The page doesn't mention tools, actions or integrations, so I'd treat it as a conversation model for now.
When can I try Dessa AI?
You can hear the demos now, while the download and waitlist links were marked as coming soon on 24 October 2026.
If you'd like my team to help you pick the right AI tools for your business, book a free SEO strategy session and we'll go through it together.
📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉
🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉
About Julian
I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
- I've built a 7-figure SEO and link-building agency, Goldie Agency, with 70+ people in Bangkok.
- I've grown a YouTube channel to 400,000+ subscribers.
- I wrote the book "Link Building Mastery", which covers nine link building strategies.
- I run GoldieBench, where I score AI models on real one-shot builds.
→ Get my best AI training inside the AI Profit Boardroom
My early verdict stands until I can test it myself, which is that the conversation is the best I've heard and the access is the weakest part of Dessa AI.
