Block trained its new speech model on a single Nvidia GB200 NVL72, a liquid-cooled rack that Nvidia says holds 72 Blackwell GPUs and 36 Grace CPUs. The finished model, called Dessa, runs on a MacBook Pro with an M4 Max chip or better, and everything it knows came from human conversation, with no text and no pretrained model underneath.

Block’s in-house AI lab put Dessa online Thursday with a demo site full of recorded conversations. Jack Dorsey, Block’s co-founder and chairman, promoted it on X that afternoon and added in a reply: “this is big (but also really really small…small enough to run on your macbook).”

Why it skips the text step

Most voice assistants work as a relay: one model transcribes the caller, a language model writes the answer and a text-to-speech engine reads it back. Nvidia’s open Nemotron Voice Agent blueprint is built that way. Block’s lab says that once speech is turned into text, the sarcasm and the frustrated sighs are gone, and it blames that for voice AI that feels “stilted, slow and, worst of all, annoying.”

Dessa takes audio in and sends audio out. It is full-duplex, so it keeps listening while it talks, and in one of Block’s clips it finishes a movie title the other speaker can’t recall. The lab says it learned language from “hundreds of times fewer words” than large language models need.

Nvidia went the other direction the same day. It released Nemotron 3.5 Bolt E4B, an open 4GB model for local voice assistants whose Hugging Face card lists inputs and outputs as text. It is built to be the middle link of the relay Dessa does without.

Where the team came from

The lab was founded by former members of Dessa, the Toronto deep learning startup Square bought in February 2020. When Square announced that deal, it said many people knew the startup for its deepfake detection work. Ragavan Thurairatnam, who leads the new lab and says in his X profile that he sold Dessa, wrote on Thursday: “This is a GPT3 moment for speech.”

Nobody outside Block can use the model yet. The lab is giving a small group of AI safety researchers early access to look for risks the team hasn’t anticipated, and everyone else gets a waitlist. Cloned voices are already part of the impersonation campaigns hitting U.S. businesses that WebProNews reported on this month. Block builds its own safety argument around listening, saying AI is “still nearly deaf to the emotion in our voices.”

The project page gives no parameter count, benchmark scores or release date. Its only hint at timing is one line: “Launch links are coming soon.”