Block Dessa Voice Model Runs Locally on a MacBook M4 Max

7 min read · Last updated: October 24, 2026

Block's Dessa voice model is a full-duplex speech-to-speech model from Jack Dorsey's company. It listens and talks at the same time, and it laughs, pauses and sounds tired when it says it is tired. Block says it was trained entirely on human speech using a single NVIDIA GPU, and that it runs locally on a MacBook M4 Max or better. There is no public access yet.

Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.

The report

• TechCrunch: Block's new voice model sounds eerily human, by Ivan Mehta, October 24, 2026

The announcements (quoted inside that report)

• Jack Dorsey on X, October 23, 2026
• Ragavan Thurairatnam on X, October 23, 2026
• Block (@blocks) on X, October 23, 2026, plus Block's own blog post

What happened with Block's Dessa voice model?

On October 23, 2026, Block, the payments company behind Square, released a voice model called Dessa from its research lab. Researchers posted demo clips where the model laughs, pauses, sounds tired after saying it was tired from work, nearly breaks down emotionally, and plays a guessing game with a Block team member named Cara.

The detail developers should care about is where it runs. According to TechCrunch's October 24, 2026 report, Block said Dessa runs locally on a consumer-grade computer such as a MacBook M4 Max or better. You cannot try it yet. Block has not opened public access and has instead put out a call for safety researchers to test it.

Key takeaways

• Dessa is full-duplex: it can talk and listen at the same time, like OpenAI's GPT-Live model released earlier in 2026.
• Block says it was trained fully on human speech and on a single NVIDIA GPU.
• It runs locally on a MacBook M4 Max or better, according to Block.
• There is no text in the pipeline driving the speech, which Block says is why current voice AI sounds unnatural.
• Access is closed for now. Block has asked safety researchers to test it first.

What makes Dessa different from other voice AI?

Most voice assistants convert your speech to text, send that text to a large language model, then turn the reply back into speech. Block says Dessa skips the text step entirely. It is speech in, speech out, and its knowledge comes from the conversational data it was trained on rather than from a paired language model.

Block's blog post, quoted by TechCrunch, puts the problem bluntly:

"Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can't carry. The most powerful AIs in the world can't hear how you feel."

The company also called the text-first approach "potentially dangerous" and said it leads to a worse experience.

Here is how the two approaches compare, based on what the report describes.

	Typical voice AI	Block's Dessa (as claimed)
Pipeline	Speech, then text, then speech	Speech to speech
Where knowledge comes from	A paired large language model	The conversational data it was trained on
How emotion is produced	Voice tags in text (e.g. ElevenLabs v4, Fish Audio Drama 3)	Built into the model
Listening while talking	Varies	Yes, full-duplex
Training data	Not covered in the source	Human speech only
Runs on	Not covered in the source	MacBook M4 Max or better, consumer GPUs
Public access	Available products	None yet

One caveat from TechCrunch: it is not clear whether Dessa's emotion is being steered by some kind of tags. Block claims it is built in. Nobody outside Block can check that yet.

What does full-duplex mean in a voice model?

Full-duplex means the model can listen and speak at the same moment, the way two people on a phone call can. A half-duplex system waits for you to finish, then replies. Full-duplex lets the model react while you are still talking, which is what makes interruptions, quick laughs and overlapping speech feel natural.

TechCrunch compares Dessa directly to GPT-Live, the full-duplex model OpenAI released earlier this year. So Dessa is not the first full-duplex model. Block's claim is about quality and training: Jack Dorsey called it "the most authentic full-duplex speech model ever built."

Ragavan Thurairatnam, posting on October 23, 2026, went further. He said it is the first AI model you can have a realistic conversation with "that was trained entirely on speech," and called it "a GPT3 moment for speech."

Treat those as the company's claims. The demos are company-picked clips.

Can you run Dessa locally on a MacBook?

Block says yes, on a MacBook M4 Max or better, and Thurairatnam said it runs on consumer GPUs. But you cannot download it today. There is no public access, and Block has not published hardware requirements beyond that example in the report. Until weights or an app ship, "runs locally" is a claim, not something you can test.

Still, the claim matters. A voice model that fits on a laptop changes three things:

1. Latency. No round trip to a server for every sentence.
2. Cost. No per-minute API bill for the audio.
3. Privacy. The audio of the conversation does not have to leave the device.

TechCrunch makes the cost point explicitly. Square already uses voice AI to handle ordering. With its own on-device model, Block could power more business and consumer experiences at a lower cost.

If you want to see what local AI looks like today with models you can actually download, our guide to running Google Gemma 4 locally is the practical starting point.

Why does this matter to a developer?

It matters because it attacks the architecture most voice apps are built on. If speech-to-speech models that run on a laptop get good, the standard stack of transcription, a language model and text-to-speech starts to look like the slow, lossy option. Your voice product's quality ceiling would move from the language model to the speech model.

I think the most interesting line in the whole report is the one about knowledge. Block says Dessa's knowledge stems from its conversational data, not from a large language model. That is a trade-off, and it cuts both ways.

• Good for: feel, timing, emotion, small talk, games, companionship-style conversation.
• Open question: anything that needs accurate facts, tools or your business data. The report does not say how Dessa would look up an order, a price or a policy.

So for now, read Dessa as a signal about where voice is going, not as a drop-in replacement for your current stack.

Why is Block asking safety researchers to test it first?

Block has not released Dessa publicly and has put out a call for safety researchers to test the model. TechCrunch says this is likely to test cases where a voice agent could trick the person on the other end. A model that sounds tired, emotional and human on demand is exactly the kind of tool that makes impersonation easier.

That is the uncomfortable side of the demos. The clip where the bot sounds like it is almost breaking down is impressive. It is also the reason a caller might not realise they are talking to software.

If you build anything voice-based, this is your problem too, not only Block's. Your users will soon assume that a human-sounding voice can be a machine. I wrote about the gap between AI demos and AI products in AI is not ready: why we are being sold an unfinished product, and voice is the sharpest example of it yet.

What changes in practice

Nothing changes in your codebase this week, because you cannot call Dessa. What changes is how you should plan.

If you are building	What to rethink
A voice support or ordering bot	How much of your quality loss comes from the text step
An on-device or offline app	Whether a laptop-class speech model fits your target hardware
A product with per-minute voice API costs	What your margin looks like if local models become viable
Anything that calls real people	How you disclose that the caller is an AI

If you are designing a conversational product right now, the fundamentals still apply. Our breakdown of how to build a conversational AI chatbot that actually converts covers the parts that do not depend on which model you pick.

What to do this week

1. Watch the demos critically. Note what they show (emotion, timing, a guessing game) and what they do not (factual lookups, tool use, business data).
2. Map your current voice pipeline. Write down each step: speech to text, language model, text to speech. Mark where tone and emotion get lost.
3. Check your hardware assumptions. If Block's MacBook M4 Max claim holds, ask whether your users have that class of machine.
4. Add an AI disclosure line to any voice flow. Do it now, before human-sounding voices become normal.
5. Follow Block's safety research call. Public access will likely depend on what those researchers find.

FAQ

What is Block's Dessa?

Dessa is a full-duplex speech-to-speech voice model released by Jack Dorsey's company Block on October 23, 2026. It can talk and listen at the same time and expresses laughs, pauses and other emotions. Block says it was trained entirely on human speech using a single NVIDIA GPU.

Can I use Dessa right now?

No. As of October 24, 2026 there is no public access to Dessa. Block has put out a call for safety researchers to test the model first, likely to probe how a human-sounding voice agent could trick people.

What hardware does Dessa need?

Block said Dessa can run locally on a consumer-grade computer such as a MacBook M4 Max or better. Block researcher Ragavan Thurairatnam also said it runs on consumer GPUs. No fuller hardware spec appears in the report.

How is Dessa different from ElevenLabs?

According to TechCrunch, models like ElevenLabs v4 and Fish Audio Drama 3 use voice tags to drive emotion in speech generation. Block claims emotion is built into Dessa and that there is no text in its pipeline. TechCrunch notes it is not yet clear whether Dessa's emotion is steered by some kind of tags.

Map your own voice pipeline this week and mark where the emotion gets lost, then grab the free tools on Art of Code for the rest of your build. If you want a model you can run on your own machine today, start with our Gemma 4 local setup guide.

Written by Adithya Guttha, Founder of Art of Code.