# Block Dessa, explained

> Block Dessa is a full-duplex speech-to-speech voice model from a research lab inside Block (the company behind Square and Cash App), announced October 23, 2026. It takes audio in and gives audio out, with no text step and no text model underneath. It is a research preview: early access goes to a small group of AI safety researchers, and everyone else joins a public waitlist. Block has published no benchmarks, parameter count, API, price, or weights. Independent explainer, not affiliated with Block.

## Key facts

- Announced: October 23, 2026, by Block's research lab (led by Ragavan Thurairatnam, reporting to Jack Dorsey).
- Type: full-duplex speech-to-speech model. It listens and talks at the same time, so it can interrupt and backchannel.
- Training: Block says it was trained from scratch on human conversation audio only, with no text and no pretrained models, on a single NVIDIA GB200 NVL72 rack.
- Runs on: Block says a MacBook Pro M4 Max or better, with inference faster than real time. No latency figures published.
- Access: research preview. AI safety researchers apply with an institutional email. Public waitlist for everyone else. Launch links coming soon, no date announced.
- Not published: parameter count, benchmarks, API, price, weights.
- Demos: recorded conversations only. Dessa laughs mid-sentence, raises its volume when asked, and finishes a speaker's sentence. One demo pairs Dessa with OpenAI's GPT-Live-1, which is not affiliated with Block and has not endorsed Dessa.
- Caveat: Block's "most authentic" claim has no published metric. Full-duplex voice itself is not new: Kyutai's Moshi did it in 2024, but from a text model.

## Links

- [Block Dessa, explained](https://lab.awaara.cn/block-dessa/): full guide, access checker, comparison, and FAQ.
- [Block's Dessa page](https://block.xyz/dessa): official announcement, demos, waitlist, and researcher application.

## Optional

- This page's access checker only sorts visitors into the access groups Block has announced. It cannot grant access.