Block Introduces Dessa Speech Model in Research Preview

Block’s new speech model, Dessa, can receive and generate audio during the same conversation without waiting for the speaker to finish, according to the company. Introduced in an October 23, 2026 announcement, the model was available as a research preview with access through a waitlist at the time of The Indian Express report.

The company describes that conversational operation as full-duplex. In Block’s account, Dessa processes incoming speech while producing its own audio response, rather than handling listening and speaking as separate turns. Block also says the process does not require an intermediate transcription of the incoming audio.

Public demonstrations provided the available evidence of those capabilities, according to The Indian Express. At the time of its report, there was no published evaluation against which to assess the model. Those examples alone do not establish whether Dessa delivers consistent performance in conversations beyond the demonstrations.

For local use, Block says Dessa can operate on a MacBook Pro equipped with an M4 Max chip or better. Independent testing of that hardware assertion was not established in the reporting.

Block says human conversations supplied the speech data for training, with neither text-based training material nor an existing pretrained model used as a starting point. Its research team also said a single NVIDIA GB200 NVL72 system was used for the training, according to The Indian Express. That report did not independently verify the team’s account of the training hardware.

Access remained at the preview stage described in the report. The Indian Express also stated that Block had not published pricing, an API, downloadable model weights or the model’s parameter count at that time.

* This article is based on publicly available information at the time of writing.