Block's in-house AI lab unveiled Dessa on October 23, 2026, a full-duplex speech-to-speech model trained entirely on human conversation audio, with no text intermediate, no API, no benchmark scores, and no public release date. The pitch is architectural, not commercial: every voice assistant currently on the market stitches together three stages (speech-to-text, a language model, text-to-speech), and Dessa doesn't.

That cascade is where the interesting information dies. As The Decoder notes, converting audio to text first strips out sarcasm, laughter, and the pitch drop that signals a frustrated caller. Dessa keeps the audio end-to-end. In the clips Block has released, AI Chat Daily describes the model laughing mid-sentence, raising its volume on request, and gasping at an anniversary reveal.

If that generalizes, it reorders the voice-agent vendor landscape. Block already sells to small businesses through Square, which uses voice AI for ordering, so there's at least a plausible product distribution path from research preview to the booking lines and support flows a 5-to-30-person operation actually runs. The near-term reference point in cascaded systems is Microsoft's MAI-Transcribe-2 streaming debut at 2.5% WER and 0.13s latency, which is state of the art for the old stack. Dessa isn't competing on those numbers. It's arguing the numbers measure the wrong thing.

The claims around how it was built are, on paper, aggressive. Block says Dessa was trained on a single NVIDIA GB200 NVL72 rack (72 Blackwell GPUs), runs faster than real time on a MacBook Pro with an M4 Max chip or better, and used "hundreds of times fewer words" than large language models. None of that comes with the context that would make it checkable: no stated training duration, no hours of audio, no model size, no data budget behind the "fewer words" claim.

And then there's the evidence problem. OrcaRouter counts more than a dozen conversation recordings and zero benchmark numbers. The clips are described as live and unsteered, with subtitles added afterward. There's no held-out test set, no end-to-end latency figure in milliseconds, no intelligibility-under-noise score. It's a curated reel.

That's the posture to read Dessa through. Full-duplex speech-native models are coming for the cascaded stack; the question for anyone sizing a voice-agent contract this quarter isn't whether Block ships Dessa, but how long the vendors who sound dated in Block's demos have before their own customers notice.

Research previews with no benchmarks and polished demo reels have a lineage. The pattern is familiar: the architectural claim is the product, and the product follows the discourse by a year.

Block's Dessa Skips Text Entirely, Listens and Talks at Once, and Runs on a MacBook

Sources