Dessa is Block’s research-preview speech-to-speech model: full-duplex voice that listens and talks at the same time, trained—per Block—from scratch on human conversation audio only (no text pipeline, no pretrained text models in the story they publish). Official page: block.xyz/dessa. This article is an evergreen product/model brief for Articles. For the launch-day News brief on the same sources, search AI Rank News for slug block-dessa-speech-model-1023 (separate from this evergreen Articles guide).
What Block Publishes
From block.xyz/dessa (checked October 25, 2026, America/Los_Angeles), with press corroboration:
- Category: full-duplex conversation model; Block calls it trained only on speech / human conversation (“Sounds! No text, no pretrained models” on the release page).
- Behavior claims on the page: hears and speaks emotion, sarcasm, laughter, sighs; inference “faster than real time”; demos are live person↔Dessa recordings with subtitles added later.
- Local run claim: “runs locally on a MacBook Pro M4 Max or better.” Block has not published a public parameter count, VRAM floor, or supported-chip matrix beyond that line—treat those as not disclosed.
- Training hardware named on the page / press: a single NVIDIA GB200 NVL72 (also reported by TechCrunch and The Decoder).
- Motivation (Block): curiosity, better human–AI understanding, and safety (voice AI that misses emotion can be dangerous in high-stakes moments—Block’s framing).
- Early access: “Calling all safety researchers” — a small group of AI safety researchers get early access via the Request access application on the same page. A general “Join the waitlist” control also appears; independent trackers have noted the public waitlist may still be non-functional depending on browser/date—rely on the live page, not third-party summaries.
TechCrunch (Oct 24, 2026), The Verge (Oct 23, 2026), and The Decoder (Oct 24, 2026) match the speech-only / full-duplex / M4 Max / safety-researcher framing. TechCrunch also notes no public access for general users at publish time and contrasts Dessa with text-tag emotion systems (e.g. naming ElevenLabs / Fish Audio as examples of tag-driven emotion elsewhere—not as Block’s own API list).
Weights, API, Price
As of the checked Block page and the same press set:
- No public weights package, model card with downloadable checkpoints, or open-source license for the weights is published on block.xyz/dessa.
- No public API docs or self-serve API key product for Dessa appear on that page.
- No public price for Dessa access, hosting, or inference.
Third-party field notes such as getdessa.com/faq (operator Jianming Duan; not Block) reach the same “unconfirmed / not published” conclusion for weights, API, and price—useful as a cross-check, not as an official Block document.
Do not assume a waitlist click or demo clip equals permission to host or redistribute the model.
Who It Sits Next To (High Level)
These are different products with different access models—not a scoreboard:
| Name | What public materials emphasize | Access pattern (high level) |
|---|---|---|
| Dessa (Block) | Speech-only train; full-duplex; local MacBook M4 Max+ claim; safety-researcher early access | Research preview; apply on block.xyz/dessa |
| OpenAI GPT-Live / Advanced Voice | Full-duplex / richer voice modes in ChatGPT stack (press comparisons) | Consumer / API product surfaces OpenAI ships |
| Hume EVI, Kyutai Moshi, others | Empathic or duplex voice research/products (Decoder’s roundup) | Each project’s own site |
Block’s own demos include a clip of Dessa talking with GPT-Live-1. That is a published demo pairing, not a published leaderboard.
How to Apply (Official Only)
- Open block.xyz/dessa.
- For researcher early access, use Request access / the research application form on that page (privacy notice dated October 21, 2026 on the same page).
- Ignore lookalike coins, invite codes, or third-party “Dessa access” listings unless Block links them from block.xyz.
Official Links
- Block release page: block.xyz/dessa
- TechCrunch: Block’s new voice model…
- The Verge: Jack Dorsey’s Block made an AI model…
- The Decoder: Block’s new AI model Dessa…
Sources
- block.xyz/dessa
- TechCrunch, 2026-10-24
- The Verge, 2026-10-23
- The Decoder, 2026-10-24
- getdessa.com/faq (independent field notes; not Block)
Keep exploring: see which open-source AI projects are rising today in the open-source rankings on AI Rank.
Or 📱 get the AI Rank app to follow them on your phone.
Prepared with automated editorial assistance and checked against block.xyz/dessa, TechCrunch, The Verge, The Decoder, and getdessa.com/faq on October 25, 2026 (America/Los_Angeles). No Dessa weights or API were obtained for this article.