Block said on October 21, 2026 that early access to Dessa, its new speech-to-speech conversation model, will go only to a small group of AI safety researchers.
On the Dessa launch page, the company describes Dessa as a full-duplex, speech-native system that takes in audio and replies in audio rather than routing through a text model. Block also says it trained Dessa from scratch on human conversation only, with no text data and no pretrained models.
The focus is live conversation. Block says Dessa can listen while speaking, handle interruptions, and produce backchannel cues during an exchange. The company also claims inference runs faster than real time and that the model can run locally on a MacBook Pro M4 Max or better.
There is still no confirmed public API, pricing, model download, or general rollout. For now, Dessa looks more like a research signal for speech-first agents than a tool teams can deploy today.