Block introduced Dessa, a speech-to-speech model from its research lab, on 23 October 2026. The company says its Dessa research page shows a model trained from human conversation without text or a pretrained model. TechCrunch reported the next day that Dessa had no public access. Block is inviting a small group of AI safety researchers to request early access, while its launch links say they are coming soon. For developers, this is a research preview rather than a model they can integrate today.
What Block says Dessa can do
Dessa takes audio in and produces audio out, according to Block. The company calls it full duplex: it can listen while it speaks, so someone can interrupt it and it can offer short listener responses, known as backchannels. Block presents recordings as live conversations with the model and says nobody steered them. It added subtitles afterward. These selected examples show what the team chose to demonstrate, not how reliably the behavior holds across speakers, languages, longer sessions or background noise.
Block also says it trained Dessa on one NVIDIA GB200 NVL72 rack and that inference runs faster than real time on a MacBook Pro with M4 Max or better. NVIDIA describes the GB200 NVL72 as a 72-GPU rack, not a single graphics card. Block has not supplied measured latency, memory requirements, a parameter count or a reproducible evaluation on its public page. The hardware and speed statements are company claims rather than independently verified deployment results.
What developers can verify today
Block’s page offers recordings and an early-access request for safety researchers. It does not provide public model weights, API documentation, a model card or a license in the material reviewed on 25 October 2026. Its local-running claim therefore does not yet amount to an installation path for other developers.
Before teams can assess Dessa for an application, they would need release terms, technical documentation and tests that disclose latency, speech quality, language coverage and behavior under interruptions and noise. The public demos raise useful questions about speech-only training and overlapping conversation. They cannot answer how the model performs outside Block’s chosen examples.
