Block’s Dessa Learned to Talk From Speech Alone and Runs on a MacBook Pro, but Access Is Waitlist Only
Block's Dessa is a full-duplex speech model trained only on human conversation, running locally on an M4 Max MacBook Pro. Access is waitlist only for now.
Block has built a voice model that never read a word, and it wants you to hear the difference. Dessa, announced on 24 October by the company’s small AI research lab, is a full-duplex speech-to-speech model trained from scratch on nothing but recorded human conversation. Block says it runs locally on a MacBook Pro with an M4 Max chip or better, and that inference is faster than real time.
The claim that matters is not the voice quality, although the demos are unsettlingly good. It is the training method. Most voice assistants turn your speech into text, think in text, then turn text back into speech, which throws away tone, pauses and emotion along the way. Dessa skips that round trip entirely.
Block is not selling access yet. The page offers a waitlist, a download link marked coming soon, and an application form for AI safety researchers who want early access.
Quick answer: Dessa is a speech-native, full-duplex model from Block. It learns from audio only, runs on a MacBook Pro M4 Max or better, and is not open source or generally available. Block says selected safety researchers get early access while everyone else joins a waitlist.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. https://t.co/r0qFLdvd6c
— jack (@jack) October 23, 2026
What Dessa actually is
Full duplex means the model listens and speaks at the same time. In practice that shows up as backchannel sounds, a quick “yeah” while you are still talking, and interruptions that land at the right moment instead of after a fixed delay. Block’s demo recordings are presented as unscripted conversations in which the model picks its own words, timing and tone.
The training claim is the unusual part. Block says Dessa used no text, no transcriptions of its audio and no pretrained model as a starting point. Everything it knows about language, and about the sound of conversation, came from speech. The company argues that text flattens meaning: stress the word “she” or “stole” in the same sentence and you get a different accusation, while text storage loses the difference.
Block also says the model learned language from hundreds of times fewer words than a large language model needs, and that the whole training run fit on a single NVIDIA GB200 NVL72 rack. That is one machine, not a data centre campus, which is the kind of detail that makes researchers sit up.
Inference runs on the device. Block names a MacBook Pro with an M4 Max or newer as the minimum, which puts the model on a laptop rather than behind an API. No audio leaves the machine, and there is no network round trip between your sentence and the reply.
How it differs from the voice assistants you already use
| What changes | Typical text-based voice assistant | Dessa |
|---|---|---|
| Input | Speech converted to text first | Audio in, audio out |
| What the model hears | Words and punctuation | Words plus tone, pauses, sighs, laughter and accent |
| Turn taking | Waits for you to stop speaking | Listens and speaks at once, can backchannel |
| Where it runs | Usually a cloud service | Locally on an M4 Max MacBook Pro or better |
| Latency | Network plus model time | Block says faster than real time on device |
That table explains why the demo clips sound different from a smart speaker. The interesting engineering is not that the voice is convincing. It is that the model kept the information a transcription step normally deletes, and then learned to act on it.
The hardware claim is the real story for developers
Voice models have been improving for years, but most of that progress arrived through hosted APIs. Running a conversational speech model on a laptop changes who can build with it. A clinic in Pune, a call centre in Manila or a solo developer in Toronto could ship an offline voice agent without sending customer audio to a third party, which matters for regulated work and for anyone paying per minute.
The trade-off is hardware. An M4 Max MacBook Pro is a serious purchase, and Block has not published memory use, context limits, supported languages or a quantised build. Until those numbers exist, the local claim is a promise about a machine most developers do not own yet. Compare that with the open weight releases covered on this site, such as Kimi K3.1 Pro, where the weights are promised but the hardware bar is far higher.
What is confirmed, and what is not
| Claim | Status |
|---|---|
| Trained only on human speech, with no text and no pretrained model | Block’s claim, no independent audit published |
| Full-duplex audio in, audio out | Supported by Block’s demo recordings, not independently measured |
| Runs locally on a MacBook Pro M4 Max or better | Block’s claim, no public build to test |
| Trained on a single NVIDIA GB200 NVL72 | Block’s claim |
| General availability | Not available. Waitlist open, launch links marked coming soon |
| Model weights or licence | Not disclosed |
| Early access for AI safety researchers | Confirmed application form on Block’s site |
Block has published a detailed technical narrative, sample audio and a team list, which is more than many launch pages offer. It has not published a paper, a model card or benchmark scores, so the architectural claims cannot be checked from outside.
Why Block is framing this as a safety project
Calling all AI safety researchers.
We’re inviting a small group to explore Dessa, our model trained on speech alone, and help us understand what it can and can’t do.
We’d love to hear from you: https://t.co/FDmoXNPpIh pic.twitter.com/UdtlvOmSn7
— Block (@blocks) October 23, 2026
The pitch is that emotion is a safety input, not a nice-to-have. A voice agent that cannot hear distress keeps talking through a medical emergency, a scam attempt or a customer in tears. Block’s announcement says current voice AI is “nearly deaf” to that signal because text strips it out, and that Dessa is a step toward fixing it.
There is a second safety story that Block does not lead with. A speech model this convincing, available on a laptop, is also a tool for impersonation. Fraud units in India already warn about AI-cloned voices in digital arrest scams, and a local model with no watermarking requirement makes that harder to police. Block says it is inviting outside researchers to find the risks it missed, which is the right instinct and also an admission that it does not know them all.
What it means for Indian developers
The practical angle for teams in India is privacy and cost. An on-device speech model removes the per-minute API bill and keeps audio inside the device, which suits healthcare, banking and education products where customer calls cannot leave the country. It also fits the reality of patchy connectivity, since a local model keeps working when the network does not.
Two unknowns matter before anyone plans a product. First, language coverage. Block has not said how much Hindi, Tamil, Bengali or code-mixed Hinglish speech went into training, and a speech-only model with thin multilingual data will struggle where a text model would not. Second, hardware. Developer laptops in India skew toward Windows and mid-range Macs, so an M4 Max requirement limits the early audience.
The sensible move is to treat Dessa as a signal about where voice AI is going rather than a component to build on today. Watch the safety researcher reports, then the language list, then the pricing, in that order. The same discipline applies to inference efficiency claims such as DeepSeek’s DeepInfer release: the demo is the easy part, the deployment details decide whether it works for you.
Frequently asked questions
What is Dessa from Block?
Dessa is a full-duplex speech-to-speech AI model built by a small research lab at Block, the company led by Jack Dorsey. It was trained from scratch on recorded human conversation rather than text, and it is designed to hold a live spoken conversation on a local device.
Can I download Dessa today?
No. Block’s page offers a waitlist and says download links are coming soon. A small group of AI safety researchers is being given early access through an application form.
Does Dessa use text at all?
According to Block, no. The company says the model used no text data, no transcriptions and no pretrained model, and learned language entirely from audio.
What hardware does Dessa need?
Block lists a MacBook Pro with an M4 Max chip or better. It has not published memory requirements, supported languages or any other hardware configurations.
Is Dessa open source?
Block has not said. No weights, licence or model card have been released, so Dessa is currently a closed research preview rather than an open model.
Sources
Block, Dessa announcement page, 24 October 2026. Jack Dorsey, announcement post on X, 24 October 2026. Block, call for safety researchers on X, 24 October 2026.