← Back to Voice models
Block
Dessa
Released Oct 23, 2026
A full-duplex speech-to-speech model from Block's AI research lab, trained from scratch only on recorded human conversation, with no text and no pretrained models. It listens while it talks and runs locally on a high-end Mac.
- Price per minute
- —
- Pricing
- No pricing published.
- Inputs
- audio
- Tool calling
- —
- Languages
- —
- Voices
- —
- Big Bench Audio
- —
- Access
- Waitlist
- Licence
- Proprietary
Strengths & weaknesses
Strengths
- Full duplex by design: backchannels and decides when to cut in
- Trained only on human conversation audio, on a single NVIDIA GB200 NVL72
- Runs locally on a MacBook Pro M4 Max or better
Weaknesses
- Waitlist only, with early access for a small group of AI safety researchers; no API or price
- No benchmarks, latency figures or model size published
- No text in training means no stated tool use; fact learning and new languages are listed as open questions
Not notable: not on AA, and Block isn't a frontier-lab company. Moved from models.json; re-verified 2026-11-06 against block.xyz/dessa, which states the facts above but gives no publication date (only a privacy-notice effective date of 2026-10-21). releaseDate 2026-10-23 is carried over from the existing entry and news item (the announcement and Jack Dorsey's X post); it isn't printed on Block's page. The page lists no tool use, so toolCalling is null.