Block’s Dessa talks and listens at once, and runs on a laptop

Block, the company led by Jack Dorsey, has unveiled Dessa, a speech-to-speech model it says was trained from scratch on nothing but human conversation and runs locally on a MacBook Pro M4 Max or better. Access is by waitlist for now.

The details below come from Block’s own Dessa page. The Verge also covered the launch, as Techmeme’s summary of its report shows.

A full-duplex model with no text in training

Block’s small AI research lab built Dessa. The page calls it the most authentic full-duplex speech model ever built. Full-duplex means it can listen and speak at the same time, so it can drop in short reactions such as “yeah!” or “got it!” while the other person is still talking, instead of waiting for a turn.

Block says Dessa was trained with no text and no pretrained models. Most voice assistants chain speech recognition, a language model and speech synthesis. Block argues that such pipelines lose emotion, sarcasm, laughter and sighs on the way through, and says Dessa handles them natively. It also says the model learned language from hundreds of times fewer words than large language models see. As far as Block knows, it describes Dessa as the first AI you can hold a realistic live conversation with that was trained only on human speech.

On compute, Block says training used only one NVIDIA GB200 NVL72. Inference, it says, runs faster than real time, so there is no awkward wait. The page does not give a parameter count, training hours, a license or any benchmark numbers.

Block published a long list of demo recordings, from a conversation about the release to life advice, a singing sample and a session with a third-party voice AI called GPT-Live-1. The page says they are live recordings with subtitles added afterward.

Waitlist first, safety researchers earliest

Dessa is not yet available to the public. Block lists a waitlist, says “launch links are coming soon” and offers early access to a small group of AI safety researchers, who must apply with an institutional email. It names no price, no release date and no open weights.

The team behind it is led by Ragavan Thurairatnam and, per the page, was founded by former members of Dessa, the Toronto deep learning startup. It reports directly to Dorsey. Block’s page frames safety as a motive, saying Dessa is a step toward fixing a problem it describes in its text.

Who can actually use it, and what to check

For developers and researchers, the practical news is the hardware claim. According to third-party guides, a GB200 NVL72 is a liquid-cooled rack that links 72 Blackwell GPUs and 36 Grace CPUs, per Spheron’s guide. Training on a single one is a very large budget by lab standards, though far from a hyperscale cluster. Running the result on a laptop is a different matter: coverage of Apple’s 2024 launch, such as Thurrott’s report, says the M4 Max tops out at 128GB of unified memory. Block does not say how much memory Dessa needs, so the minimum configuration is unknown.

For everyday users, nothing changes yet. There is no app, no pricing and no regional availability. A local model would keep audio on the device, but Block has not described how Dessa would be offered, and its linked privacy notice, effective October 21, 2026, says it collects identification, device and usage data and does not sell user information.

For voice-AI builders, the claim to test is conversation quality without a text backbone. Block’s “world’s first” statement is its own and is hedged with “as far as we know”. No independent evaluation is cited, and the demos are curated by the company. Until researchers get access, claims about emotion handling, interruption behavior and faster-than-real-time speed remain unverified.

The signal to watch is the launch links Block promises: whether weights or an API appear, what minimum Mac is really required, and whether safety researchers publish findings on how the model behaves in long, open conversations.