Block launches Dessa: It can learn to chat just by listening to human speech, and runs natively on MacBook.

Beating AI Express: Block, the company founded by Jack Dorsey, has unveiled a voice model called Dessa. Its most unique trait is that it was trained entirely from scratch using only human conversation recordings, with no text data or pre-trained language models involved. Dessa directly receives and generates audio, supporting full-duplex conversation—meaning it can listen and speak simultaneously. It can naturally interject before the other party finishes speaking, and even laughs out loud when it catches a joke. In an official demo, when a person suddenly forgot a movie title, Dessa volunteered to fill in the blank; upon hearing an anniversary surprise, it let out an exclamation and responded with a tearful tone. Block stated that Dessa was trained using just one NVIDIA GB200 NVL72 system, equivalent to a full rack of 72 Blackwell GPUs. The model’s inference speed exceeds real-time playback, and it can run natively on MacBook Pro devices equipped with M4 Max or higher chips. To date, Block has not disclosed the model’s parameter count, standard test results, or weights, and has only granted early access to a small number of AI safety researchers.