Dessa is Block's on-device full-duplex speech model, built for developers and product teams who want voice interaction that sounds human instead of robotic. Unveiled by Jack Dorsey's Block on October 23, 2026, the model comes from a small internal lab led by Ragavan Thurairatnam that reports directly to Dorsey, and it skips the usual speech-to-text-to-speech pipeline to run audio in, audio out on consumer hardware.
Core Features
- True speech-to-speech architecture: no text intermediary, so laughter, pauses, sarcasm and tone survive the conversation.
- Full-duplex listening: it talks and hears at the same time, backchanneling with 'got it' or a laugh and letting you interrupt naturally.
- Runs locally on a MacBook Pro with an M4 Max chip or better, faster than real time, with no cloud round-trip.
- Trained from scratch on human conversation only, using no written text and no pretrained base model, on a single NVIDIA GB200 NVL72 rack.
Use Cases
- Privacy-first voice assistants and offline customer-service prototypes that cannot send audio to a server.
- Accessibility and emotional-safety research, where detecting a distressed voice matters more than transcribing words.
- Block product experiments inside Square and Cash App, where low-latency local inference is a requirement.
Pricing
Dessa is a research preview, not a product. There is no public API, no weights release and no commercial price. Early access is limited to a small group of AI-safety researchers, with a public waitlist for everyone else.
Our Take
Best for researchers and privacy-minded builders exploring speech-native AI; the trade-off is that benchmarks, licensing and a real release date are still unproven. Explore more voice work in our AI audio tools hub.
