Block's AI research lab announced Dessa on October 23, 2026, and the claims are the company's own so far. Dessa is a speech-to-speech model: audio goes in and audio comes out, with no text step in between. It is full-duplex, so it can listen and talk at the same time, interrupt, and say short backchannels like "yeah" or "got it" while you talk.
Block says it was trained from scratch on human conversation only, with no text data and no pretrained models. It also says the model picks up emotion, laughter and sighs, that it responds faster than real time, and that it runs locally on a MacBook Pro M4 Max or better. The company calls it "the most authentic full-duplex speech model ever built," and the demos it published are selected clips.
What's not established yet: there are no independent benchmarks, no published paper I could find, and the public can't use it. Block describes it as a research preview with a waitlist, and early access goes to a small group of AI safety researchers. Full-duplex turn-taking is a real technical goal, and several research groups are working on it, so the useful test will be how Dessa does in unedited conversations, measured against those benchmarks.
Source: Unite.AI on the Dessa announcement. The announcement itself is at block.xyz/dessa.
🕒 As of October 2026
