Block's full-duplex speech-to-speech conversation model, trained from scratch on human conversation alone and able to run locally on a high-end Mac.
Last verified · 2026-10-23 · by Moe Ameen
Dessa is a full-duplex speech-to-speech conversation model from Block, the payments company co-founded by Jack Dorsey. It was built by Block's small in-house AI research lab, which reports to Dorsey and was founded partly by former members of Dessa, the Toronto deep-learning startup Square (now Block) acquired in 2020 — the model carries that name. Unveiled in October 2026, it is audio in, audio out: you talk to it and it talks back, in real time.
The distinctive claim is how it learned. Block says Dessa was trained from scratch on human conversation alone — "sounds, no text, no pretrained models" — on a single NVIDIA GB200 NVL72, picking up language from hundreds of times fewer words than large language models. Because it is full-duplex, it listens and speaks at the same time and learned from real dialogue when to cut in and when to backchannel with short acknowledgments like "yeah" or "got it." Block says it natively handles emotion, sarcasm, laughter, a frustrated sigh, and a voice cracking with sadness, rather than flattening speech into text first, with inference that runs faster than real time so replies arrive without an awkward pause.
It reportedly runs locally on a MacBook Pro M4 Max or better, and lead researcher Ragavan Thurairatnam said it can also run on consumer GPUs. At announcement, access is early: Block opened a public waitlist and offered early access to a small group of AI safety researchers through an application requiring an institutional email, with "launch links coming soon." There is no published pricing, stated open-source license, or general release date, and no independent benchmarks — the demos are audio recordings, one of them a conversation with GPT-Live-1, a third-party model Block says is unaffiliated. Treat the performance claims and the "most authentic" / "world's first" framing as the company's own until third-party tests land.
Here's the honest shape of it: Dessa makes a conversation sound human; it doesn't make anything you can post. A full-duplex model is the opposite end of the pipe from a content engine — it's where talk happens, not where a feed gets filled. But a conversation is a content format in its own right, and that's where [Kompozy](/) picks it up. Record the exchange — a Dessa session, a voice-agent transcript, a podcast, or an interview — and Kompozy treats the dialogue as source: it pulls the sharpest back-and-forth lines into Quote Graphics, drafts an interview-style Blog Article and an Email Newsletter from the transcript, and cuts the audio into vertical [Clipped Shorts](/glossary/clipped-short) auto-captioned for TikTok, Reels, and Shorts. Your [Persona Brief](/glossary/persona-brief) governs every one of those so the written voice stays consistent no matter how loose the spoken conversation was.
Then Kompozy does the part a speech model structurally can't: it generates net-new branded formats and ships them. The key answer from a conversation becomes a face-locked [Persona Shorts](/glossary/persona-shorts) clip re-presented by a recurring AI-influencer avatar with its own built-in voice; [HyperFrames](/glossary/hyperframes) renders Carousel Posts, Photo Posts, and Quote Graphics pixel-exact to your brand; and [Autopilot](/glossary/autopilot) schedules and fans the whole set across the eight social platforms plus blog and email behind a per-post review gate. Dessa is the voice you converse with; Kompozy is the system that turns what gets said into an on-brand presence everywhere your audience actually is.
Dessa is a full-duplex speech-to-speech conversation model from Block (the payments company co-founded by Jack Dorsey), unveiled in October 2026. Block says it was trained from scratch on human conversation alone — no text, no pretrained models — so it listens and talks at the same time, backchannels, and handles emotion, laughter, and sighs. It is named after Dessa, the Toronto startup Square (now Block) acquired in 2020.
Not generally. At announcement Block opened a public waitlist and offered early access to a small group of AI safety researchers via an application requiring an institutional email, with launch links "coming soon." It reportedly runs locally on a MacBook Pro M4 Max or better. No pricing, license, or general release date was announced.
Block's pitch is that Dessa was trained only on human speech (not text), which it says lets the model carry emotion, sarcasm, and timing natively; that it is full-duplex, so it listens and talks at once; and that it runs locally on a high-end Mac with faster-than-real-time inference. These are vendor claims — no independent benchmarks were available at announcement.
No. Dessa holds a spoken conversation — it does not write posts, generate images or video, caption clips, or publish anywhere. To turn conversational audio into content, a content engine like Kompozy repurposes a recording into captioned clips, quote graphics, a blog, and a newsletter, then schedules and publishes across the eight social platforms plus blog and email.