Block, the company behind Square and Cash App, published Dessa on October 23, 2026: a full-duplex speech-to-speech model that Block says was trained from scratch on nothing but human conversation. The team claims there was no text data and no pretrained model in the pipeline. Dessa is not generally available. Early access is aimed at a small group of AI safety researchers.
Key Facts
| Details | |
|---|---|
| Company | Block (reports to Jack Dorsey) |
| Model | Dessa, speech-to-speech, full-duplex |
| Training claim | From scratch on human conversation only; no text, no pretrained models |
| Hardware (company claim) | Trained on one NVIDIA GB200 NVL72; local inference on MacBook Pro M4 Max or better |
| Access | Not public; early access for AI safety researchers who apply |
| Lead | Ragavan Thurairatnam |
| Announcement | Official page live Oct 23, 2026 (The Verge report same day, 19:32 UTC) |
Background
Block's page frames Dessa as audio in, audio out: it listens and talks at the same time, with backchannels such as "yeah" and "got it," and inference that the company says runs faster than real time on supported MacBooks. The stated motives are curiosity, a more natural human-AI interface, and safety, arguing that text pipelines drop emotion and timing that matter when a speaker is distressed.
The lab is a small research group inside Block founded by former members of Dessa, the Toronto deep learning startup. The Verge's Jay Peters (Oct 23) notes the demo conversations often sound natural but can turn awkward or uncanny, including a clip where Dessa initially mistakes the speaker for "Ronaldo." The Decoder (Oct 24) places Dessa among earlier emotional-voice efforts such as Hume EVI, Kyutai Moshi, and OpenAI's Advanced Voice Mode, while stressing that sounding emotional is not the same as stronger reasoning.
What We Don't Know Yet
- Parameter count, dataset size or hours of audio, training duration, measured latency, and any public benchmarks (none published on the official page).
- Licence, pricing, API, or public weights.
- When, if ever, Dessa opens beyond the safety-researcher cohort.
- Independent verification of the "speech only / no text" training claim.