Block has presented Dessa, a research preview of a model that takes speech in and produces speech out. Its team says it trained the model from human conversation alone, without text or a pretrained model. The Dessa page offers recorded conversations and an application form for limited researcher access, while a public model download is still pending. The Verge reported the announcement on October 23.

The big change

  • What changed: Block is showing a voice model trained on conversation audio and designed to listen while speaking. For now, outsiders can inspect its description and samples; general model access has not opened.
  • Why it matters: Timing, pauses and vocal cues are part of spoken exchange. A speech-based training and interaction path gives voice-system researchers a distinct design to examine, though Block has yet to publish a systematic evaluation of how reliably it handles those cues.
  • What to watch: Block is inviting a small group of AI safety researchers to probe risks. Their access and any disclosed evaluation would give outsiders more to judge than the current samples and company claims.

What Block says Dessa does

Block calls Dessa a full-duplex speech-to-speech model: it receives and produces audio, and the team says it can listen and talk at the same time. In conversation, Block says, it can interject or give brief acknowledgements while the other person is speaking. The company also says inference runs faster than real time and locally on a MacBook Pro with an M4 Max chip or better. These are specifications and capability claims from the research preview, not results from a BIG CHANGE test.

The team says it trained Dessa from scratch on human conversation, with no text and no pretrained model, using one NVIDIA GB200 NVL72 system. That describes its stated training input and hardware; the page does not disclose the data set, model size or a reproducible evaluation. Block argues that audio preserves cues such as laughter, pauses and tone that a text transcription can lose. Its further claims that Dessa understands emotion and makes voice AI safer require evidence beyond the published examples.

Block presents several recordings as live conversations without a person steering the model and says subtitles were added afterward. We inspected the page and its descriptions, but did not listen to the recordings or run Dessa. Those examples cannot establish how consistently the model behaves across speakers, languages or difficult conversations.

Access remains limited

The page invites a small group of AI safety researchers to request early access. Its form asks for a name, institutional email and an optional research link. A prominent “Download” control is currently a placeholder, accompanied by the page's note that launch links are coming soon. Block says local inference is possible on suitable hardware, but the page does not provide a public model download, license, API, pricing or general release date.

That distinction matters to builders evaluating voice systems today. They can examine Block's stated approach and recordings; they cannot yet reproduce its training claims or compare a released Dessa build in their own application. Our earlier coverage of Mistral's Voxtral 3 speech models and voice-agent SDK concerns a separate release, with a different documented access path. There is no evidence connecting Dessa to Block's Square voice-ordering kiosk.

Sources & further reading

  • Block's Dessa research preview, accessed October 24, 2026. The undated company page sets out its training, interaction, hardware and access claims and hosts recordings. It does not provide independent measurements or a public model download. The privacy notice on the page is effective October 21; that is not an announcement date.
  • Jay Peters, The Verge, October 23, 2026. This dated report establishes that Block announced Dessa by October 23 and describes the published samples and limited safety-researcher access. Its discussion of a recording is the reporter's account, not a BIG CHANGE listening test.