Kingy AI
Subscribe, 2.1M subscribersWork with Kingy
AI News

Dessa AI: Jack Dorsey’s Block Built a Voice Model Without Text

October 23, 2026 Blog
Dessa AI title with decorative sound waves. Original Kingy.ai editorial illustration.

Block says Dessa learned from human conversation, can listen while speaking and runs locally on a Mac. Its recordings offer a first look at the idea, with public launch links still pending.

A laugh halfway through your sentence. A quick acknowledgment while you explain something. A pause that gives you room to finish. Those small details can make a conversation feel comfortable, and they are central to Block’s pitch for Dessa.

Dessa is a speech-to-speech AI model that Block says it trained from scratch using human conversation audio alone, without text or pretrained models. The company also says it can listen and speak simultaneously and run locally on a MacBook Pro M4 Max or better.

The announcement is signed by Dessa lead Ragavan Thurairatnam. Block describes the team as a small research lab that reports directly to Jack Dorsey, its cofounder, chairman and Block Head.

That combination makes Dessa worth following. The evidence available so far consists of Block’s announcement and recordings. The page does not provide the technical disclosures or controlled results needed to establish how consistently the model performs.

Reporting note: We reviewed primary-source documentation for this article and did not have hands-on access to Dessa. Descriptions of its demos below refer to Block’s own recordings and annotations. We have not independently verified its training, hardware or performance claims. Availability was checked on October 23, 2026, Pacific time.

Start with the official Dessa demos

Block’s recordings section is the quickest way to understand what the company wants Dessa to do. It includes conversations about films, an anniversary surprise, coffee and life advice.

In the release conversation, Block marks moments where Dessa laughs and responds to a request to speak louder. In the movie conversation, the annotations highlight help recalling a title and moments when the model takes over the speaker’s turn. The anniversary clip is presented as an example of expressive reactions.

These are useful listening prompts: pay attention to when the model responds, whether it leaves room for the person and how its voice changes during the exchange. A pleasant voice alone would tell you little about its ability to manage a conversation.

Block describes these as live recordings without someone steering the model and says subtitles were added afterward. That is the company’s account of how the clips were produced. The page does not disclose how the examples were selected or what fraction of attempted conversations worked this well.

Readers can judge the published examples for themselves. Measuring reliability would require repeatable tests and access beyond those examples.

What makes Dessa’s training claim interesting?

Block’s central claim concerns how Dessa learned. According to its announcement, the model learned to converse from speech alone, without starting from an existing pretrained model.

Spoken communication gives a model more to work with than a transcript. The same sentence can sound enthusiastic, impatient or uncertain. Overlapping speech can signal agreement, a correction or an attempt to change the subject. Preserving those cues could help a voice system choose a better moment and tone for its response.

There is research precedent for tackling this directly. Kyutai’s 2024 Moshi paper describes a full-duplex speech system that models the user’s speech and its own speech in parallel. Moshi starts from a text language model and also predicts text alongside audio.

That context matters when assessing Dessa. Full-duplex conversation was already an established research direction. Block’s distinctive claim is that Dessa learned its conversational abilities without text training or pretrained models. The launch page does not provide enough methodological detail to independently reproduce that result or establish priority over every earlier project.

Full-duplex means it can listen while talking

In a full-duplex conversation, both sides can send and receive speech at the same time. For a voice assistant, that creates room for interruptions, overlapping speech and brief acknowledgments while the user continues talking.

Block says Dessa learned these behaviors from real conversations. Its examples emphasize timing as much as the content of an answer.

The practical test will be whether Dessa can distinguish a useful interruption from an unwanted one. A listener saying “right” may be encouraging the speaker to continue. Someone saying “wait, I meant Thursday” needs the system to change course. Supporting simultaneous audio makes those interactions possible; handling them well requires judgment that a few recordings cannot quantify.

Dessa specs: what Block has disclosed

The following table separates company claims from details directly visible on the launch page. It is a disclosure summary, not a benchmark scorecard. Source: Block.

Item Current disclosure Evidence status
Model type Speech-to-speech with simultaneous listening and speaking Block’s description
Training Human conversation audio; no text or pretrained models Block’s claim
Training hardware One NVIDIA GB200 NVL72 Block’s claim about its training run
Local hardware MacBook Pro M4 Max or better Block’s claim; no Kingy test
Speed Inference described as faster than real time Block’s claim; no numerical latency results on the page
Parameter count and required memory Not disclosed on the launch page Documentation gap
Public access Page says launch links are coming soon; invites selected safety researchers Observed page content
API pricing, downloadable weights and license Not provided on the launch page Documentation gap
Controlled benchmark results Not provided on the launch page Documentation gap

Block also claims Dessa learned language using hundreds of times fewer words than LLMs. The announcement does not specify a comparison protocol sufficient to turn that statement into a measured efficiency advantage.

One GB200 NVL72 means a rack containing 72 GPUs

Block’s training hardware statement needs context. NVIDIA specifies that a GB200 NVL72 contains 72 Blackwell GPUs and 36 Grace CPUs in a rack-scale system.

That specification establishes what the named hardware contains. It does not verify Block’s use of it. The announcement also leaves training duration and total cost undisclosed, so it cannot support a claim that Dessa was inexpensive to train.

Can you run Dessa on a Mac today?

Block says Dessa runs locally on a MacBook Pro M4 Max or better. The launch page does not pair that claim with a public installation guide, memory requirement or downloadable model.

For now, that makes the Mac statement an indication of the team’s intended capabilities. Readers cannot use the information on the page to reproduce the setup. We have not measured its speed, memory use or power consumption on a Mac.

Local execution could be useful for people who want responsive voice interactions on their own hardware. Whether a released Dessa application could operate fully offline, what it might transmit and how it would handle recordings would depend on the implementation and privacy documentation.

The hardware claim alone does not establish those properties. It would also be premature to buy a Mac specifically for Dessa before Block publishes a supported configuration and usable release.

How can you try Dessa?

At the time of our check, the official page said “Launch links are coming soon.” It also invited a small group of AI safety researchers to request early access.

The immediate public option is to explore the recordings. Eligible researchers can use the contact option on Block’s page. We did not verify a working general-access signup or download flow and cannot provide a public release date.

Keep the distinction between model access and media downloads clear. The page offers downloads of conversation clips; those files do not give you Dessa itself.

There is also insufficient information on the page to quote a consumer subscription price, an API rate or commercial licensing terms.

What would meaningful Dessa benchmarks measure?

A useful evaluation would measure both conversational behavior and the quality of the answers. These are proposed tests, not results from a Kingy benchmark.

Response timing should include the delay before an answer and how quickly the system reacts to an interruption. Tests should report typical performance and slower cases, with hardware and audio conditions recorded.

Turn-taking needs its own assessment. Does the model stop when corrected? Can it acknowledge a speaker without talking over them? Does background speech trigger an unwanted response? Repeating those situations across speakers would reveal more than one successful exchange.

Answer quality also matters. Natural delivery can make an incorrect statement sound persuasive. An evaluation should test factual accuracy and instruction-following alongside listener preferences, using the same tasks and conditions for any competitors.

Finally, Block’s safety rationale deserves direct testing. Recognizing distress and responding appropriately are separate capabilities. An expressive or sympathetic-sounding response cannot establish either one by itself.

For readers today, the useful next step is to hear Block’s published recordings. A practical setup guide and a defensible performance comparison will need a reproducible release, documented requirements and access for independent testing.