Block, the company behind Square and Cash App, has introduced Dessa, a new artificial intelligence model designed to hold conversations in a human-like way. Unlike most voice assistants, which depend on processing text, Dessa has reportedly learned by listening to people talk.
Dessa is described as a full-duplex speech-to-speech model, meaning it can listen and talk at the same time. Block says this allows it to interrupt, acknowledge what a speaker is saying, and reply without waiting for a turn to end.
How Dessa was trained on audio alone
Block says Dessa was built from scratch using only recordings of human conversation. It reportedly does not rely on written text or pretrained language models, so its grasp of language, facts and dialogue comes solely from sound.
The company argues this approach helps the model notice cues that conventional speech-to-text systems may miss, including sarcasm, hesitation, laughter and shifts in emotion. On its website, demos show the model laughing mid-sentence, raising its volume when asked, and completing a sentence when a speaker forgets a movie title.
In one demo, Dessa converses with OpenAI’s GPT-Live-1 voice model. Block says the recordings show live exchanges with no one guiding the responses, but the demos are first-party material and should be judged as such.
Hardware requirements and speed claims
Block says Dessa was trained on a single Nvidia GB200 NVL72 system containing 72 Blackwell GPUs. The finished model can reportedly run on a MacBook Pro with an Apple M4 Max chip or better. Block also claims it runs faster than real time, so conversations avoid noticeable pauses.
Limited access and missing disclosures
Dessa is not yet publicly available. Block is inviting a small group of AI safety researchers to test it, while others have been placed on a waitlist. The company has not shared its parameter count, model weights or standard benchmark results.
