Block's Dessa: a voice AI trained on human audio alone, one Nvidia rack, a MacBook to run it
Where it stands: Block, Jack Dorsey's payments company, has shown Dessa, a full-duplex speech-to-speech model it says was trained from scratch on human conversation alone, with no text and no pretrained models. Block says one Nvidia GB200 NVL72 rack trained it and that it runs locally on a MacBook Pro M4 Max or better. Published voice samples sound strikingly human, though some are uncanny. It isn't released: there's a waitlist, and AI safety researchers have early access. Watch for a public release, a model card or paper, and independent tests of the hardware and training claims.
Latest
- 2026-10-24 (~8:00 AM ET, Gizmodo — gizmodo.com; TechCrunch — techcrunch.com/2026): reviewers listen to the published voice samples. Gizmodo calls many of the voices human-ish but uncanny; TechCrunch calls the model "eerily human" and notes Block says Dessa's knowledge comes from its conversational training data rather than a paired language model. Still no public release or independent evaluation.
- 2026-10-23 (~3:32 PM ET, The Verge — theverge.com; Block — block.xyz/dessa): Jack Dorsey's Block announces Dessa, a full-duplex speech model, "audio in, audio out", trained "from scratch on nothing but human conversation" with no text and no pretrained models. Block's own page says one Nvidia GB200 NVL72 rack trained it, it runs faster than real time and locally on a MacBook Pro M4 Max or better, and it learned language from far fewer words than LLMs. It is built by Block's AI lab, founded by the team behind the Toronto startup Dessa. Not released: a waitlist, with early access for a small group of AI safety researchers. Block shared demo chats, some awkward or uncanny. Watch: release, model size, a model card or paper, independent tests.