Block Dessa: a full-duplex conversation model trained only on human speech, no text
What is it?
Dessa is a speech-to-speech conversation model from Block's AI research lab. It was trained from scratch on human spoken conversation only, with no text and no pretrained models. Because it's full-duplex (it listens while it talks), it can drop in a 'yeah' or 'got it' while you're speaking, and it expresses laughter, pauses, and emotion in its voice. Audio goes in and audio comes out, with no separate LLM in between. Block says it trained the model on a single NVIDIA GB200 NVL72 rack and that it runs locally on a MacBook Pro M4 Max or betterVendor claim.
Why does it matter?
Most voice agents today chain speech recognition, an LLM, and speech synthesis, or bolt voice onto a text LLM. Dessa takes a different route and learns language from speech alone. A model small enough to run on a laptop matters for on-device assistants and privacy-sensitive support calls. That said, its knowledge comes only from conversational data, and the voice is hard to tell from a human; Block itself says it wants deception risks tested first, so it's too early to build products on it.
Maturity
ResearchThere's no public API or weights, only a waitlist, plus an early-access form for a small group of AI safety researchers using institutional email. Block hasn't published benchmarks or a parameter count.
How to track it
- Access: watch whether the block.xyz/dessa waitlist opens up to the public, an API, or released weights.
- Safety: look for follow-up reports from Block's lab on how early-access safety researchers assess voice deception risks.
- Local use: check Jack Dorsey's X posts for a local build or app that backs up the 'runs on an M4 Max' claim.