Dessa AI, Block's voice model: demo, access and limits

Dessa AI is Block's voice model. It is not the musician Dessa, and it is not the DESSA student assessment used in schools.

What is Dessa AI?

Block released Dessa AI on October 23, 2026 as a research preview of a conversational speech model. Its own one-line description reads "A research preview of a full-duplex speech model." [1] Block says full-duplex means it listens and speaks at the same time [1].

It comes from a very small Block lab formed by former members of Dessa, the Toronto deep learning startup that Square (now Block) acquired on February 7, 2020 [2]. The 2026 model is a new project. It is not the 2020 product. Jack Dorsey also announced the model on X [3].

How does the Dessa voice model work?

Block says Dessa is a full-duplex system that listens and speaks at the same time [1]. It calls the model speech-to-speech, with audio in and audio out [1]. Block says it was trained from scratch on human conversations only, with no text and no pretrained models, unlike a chain of speech-to-text, a text language model and text-to-speech [1]. With no paper or weights public as of October 24, 2026, outsiders cannot check whether it uses text internally when it runs.

Block also says Dessa:

• learned turn-taking and short backchannel cues, such as yeah or got it, from real conversations [1]
• handles emotion, sarcasm, laughter, sighs and a voice cracking with sadness [1]
• was trained on one NVIDIA GB200 NVL72 and hundreds of times fewer words than language models, while noting that language models also learn programming languages and other text material [1]
• runs inference faster than real time, with no millisecond figures given [1]
• runs locally on a MacBook Pro with an M4 Max or better [1]

Jack Dorsey's reply says only that it is small enough for a MacBook [4], and Ragavan Thurairatnam's post mentions consumer GPUs [5]. This page follows Block's narrower wording.

Is there a Dessa demo?

Yes. Block's page has recorded conversations with Dessa that you can play, and Block posted three videos on X. Effect Video hosts none of this media.

Block says every clip on its Dessa page is a live, unscripted recording, with captions added afterward [1]. These are the recordings Block selected for its own page [1]. The main groups are:

• Talking about the release: Block says the model laughs in the moment and speaks louder when asked.
• Remembering a movie: Dessa finishes a thought about a film.
• An anniversary surprise: the page describes a gasp and an emotional response to a surprise.
• Expressive conversation and life advice: clips about tone, pacing, jokes, coffee, stress and empathy.
• Dessa with GPT-Live-1: two AI voices talking.
• More samples: the page lists 15 more clips as of October 24, 2026, including gaming and refunds.

Block's official X videos from October 23, 2026 include an intro, a guessing game with Dessa and an invitation to AI safety researchers.

How do I use Dessa AI?

You cannot yet. As of October 24, 2026 there is no public app, download, API, pricing or open weights for Dessa [1].

Block invites a small group of AI safety researchers to request early access with a university, company or lab email [1]. Its page says launch links are coming soon; it does not offer a public download [1].

In one demo clip's official subtitles, the voice calls itself open source AI. That is dialogue, not an open-source license. Block's page provides no weights or license for the model [1].

What is the Dessa release date?

Block released the Dessa research preview on October 23, 2026 [1]. As of October 24, 2026, Block has announced no public launch date. This speech model is unrelated to the film O'Dessa.

Dessa vs GPT Live

No public benchmark compares Dessa with GPT Live as of October 24, 2026 [11] [12]. Block's own demo is a recorded conversation between Dessa and the third-party GPT-Live-1, whose developer Block says is unaffiliated and has not endorsed Dessa [1].

Some Hacker News commenters praise the conversational flow while questioning intelligence and tool use. These are commenter impressions, not measured comparisons. No public controlled evaluation establishes that speech-only training causes those strengths or limits [14].

Block calls Dessa the most authentic full-duplex speech model built and, as far as it knows, the first trained only on human speech for real-time conversation [1]. This page does not independently verify those superlatives or rank voice models.

Who made Dessa?

A very small Block lab built Dessa. Ragavan Thurairatnam leads it, and the lab reports directly to Block co-founder and Chairman Jack Dorsey [1]. Block says the team has worked on deep learning since 2012 and on generative voice for over eight years [1].

The page does not state a headcount. It credits 12 people as of October 24, 2026: Thurairatnam as lead, six researchers (Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat, Marc Tyndel), three interns (Cara Ji, Jibran Iqbal Shah, Dvir Zagury-Grynbaum) and two former interns (Aryan Sharma, Olivia Yan) [1]. These are credits, including former interns, not a stated staff headcount.

What are the limits?

As of October 24, 2026:

• Research preview: the public cannot use Dessa. Block provides no public app, API, pricing, open weights or public launch date. A small group of AI safety researchers can request early access [1].
• Selected recordings: Block selected the demos, and says they are live recordings with subtitles added afterward. They do not establish performance beyond those conversations [1].
• Open-source dialogue: a voice in the official subtitles calls itself open source. This is no open-source license; Block's page provides no model weights or license [1].
• Training and consent: Block's public page does not disclose the training-audio sources or speaker consent. The efficiency of learning facts from speech is a research question raised by Block itself [1].
• Safety concerns: commenters raise phone scams, impersonation and emotional dependence as risks, rather than demonstrated outcomes [14] [15].
• Unchecked capabilities: Block's full-duplex, speech-only training, faster-than-real-time inference and M4 Max-or-better claims have no public benchmarks on its page. No controlled public comparison establishes a reasoning or tool-use tradeoff [1].
• Separate product: Dessa is made by Block. Effect Video provides no Dessa trials, generation, audio hosting or accounts.

How this page was put together

This page collects public material: Block's page and posts, news reports, and community threads. We have not used Dessa ourselves, and clip descriptions follow Block's own wording. Sources were checked on October 24, 2026; changing X interaction counts are omitted. If a source changes or disappears, we update or remove the line that relied on it.

For tools available here, explore AI video models or the Effect Video homepage.