On 23 October, Block, the payments company behind Square and Cash App, unveiled a speech model called Dessa. Its own page describes it as “a true speech-to-speech model”: audio in, audio out. It is a full-duplex voice model, meaning it can talk and listen at the same time. Block says it trained Dessa from scratch “on nothing but human conversation” and that it runs locally on a MacBook Pro M4 Max “or better”.
The live page is unreachable to us. block.xyz/dessa returned HTTP 403 on every attempt we made across two days. The newsroom, the site root and the www address failed the same way, and dessa.block.xyz does not resolve. The Block quotations below come from two Internet Archive captures of the page, taken on 23 and 24 October, which carry identical body copy. They are Block’s own words, quoted from the 24 October capture.
What the accounts agree on
The name, the 23 October announcement and the speech-to-speech design appear consistently across Block’s page, The Verge and TechCrunch. So does the absence of access: “there is no public access to Dessa at the moment”, TechCrunch reported on 24 October, and Block’s page offers only a waitlist plus early access for “a small group of AI safety researchers”. Block’s page publishes about 20 demo clips; both outlets link to or describe some of them. The Verge’s report is timestamped 23 October 2026, 19:32 UTC; TechCrunch’s ran the next day at 09:00 PT.
That is the verifiable outline. Almost everything a reader would use to judge the model is either the maker’s claim or a listening impression drawn from a Block-selected clip.
The data claim and the compute claim are separate
Block’s page claims the training data was speech only: “trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models.” Separately, the same page claims an unusually cheap training run: “we figured out how to train Dessa on only one NVIDIA GB200 NVL72.” TechCrunch paraphrased that second sentence as training on “a single NVIDIA GPU rack”.
These are two different claims, one about data and one about compute, and neither has been independently verified. The Verge led with the conversation-data claim; TechCrunch carried the compute figure in a sentence that also repeats the data claim. Both sentences appear in Block’s archived page, so both belong to Block; attribution is not verification.
The question Block does not answer
Block says the model learned from human conversation. It does not say whose.
The page is silent on where the conversational audio came from, whether it was recorded, licensed, purchased or collected from users, what the speakers agreed to, whether anyone could opt out, which languages are represented, and how the audio is retained or deleted. It says “No text” about training, but nothing about prompts, transcripts or text annotations used anywhere in the pipeline, so that phrase should not be stretched further than Block wrote it. The page’s privacy text is Block’s standard website notice, effective 21 October 2026, covering the data Block collects from visitors and the fields on the access form: a name, an institutional email and an optional Google Scholar or blog link. None of it addresses the training audio. The Verge and TechCrunch do not raise provenance, and no third party has commented on the record about the corpus.
A model whose only input is human speech depends on a body of recorded voices, and the terms on which those voices were obtained are the first thing a reader should be able to check. Block has published the claim and not the provenance. The precedent for why this matters is already on this site: Meta narrowed its training-data wording only after the crawl permission was questioned in its own documentation, using the same archive-based method this article relies on.
There is nothing to download and nothing to check
No model card. No benchmark, vendor-run or otherwise. No parameter count, architecture description or context length. No licence, no weights, no code. A GitHub search of Block’s organisation returns zero repositories for “dessa”, and Hugging Face carries no Block-published model. Block open-sources other AI work, including goose and Buzz, so the silence here is a deliberate difference from its usual practice. The site has also covered what an unproven vendor claim looks like when nothing follows it, in the Muse and Opus memory claim.
There is no dated release plan either. The page footer says “Launch links are coming soon”, and the only route in is a request form requiring a university, company or lab email, so even the safety-researcher channel is gated to institutional credentials. No external tester has commented on the record; every source quoted in the coverage works at Block.
One thing the page does disclose: each clip is “a live recording of a person talking to Dessa. Nobody steers it”, and clips made with OpenAI’s GPT-Live-1 carry a note that the third-party developer “has not reviewed or endorsed Dessa”. What the page does not say is how the roughly 20 clips were selected, how many takes were recorded, or whether failed runs were discarded. The Verge noted that Dessa initially mistakes a speaker for someone named “Ronaldo”; Block itself published a clip titled “Ronaldo”. The awkwardness is the company’s own. That is a claim-versus-documentation discipline the site has applied before, in the Flock gunshot case.
What “runs on a laptop” leaves unmeasured
Block’s page says inference “runs faster than real time, so there’s no awkward wait”. TechCrunch calls the target hardware “a consumer-grade computer” and quotes the Dessa lead, Ragavan Thurairatnam, describing a model “you can run on consumer GPUs”.
No size, memory footprint, quantisation level, stream count, offline behaviour or latency figure has been published, and “faster than real time” is not tied to the M4 Max or to any measured setup. The only concrete number in the entire release sits on the training side: Block says it used one NVIDIA GB200 NVL72 rack to train a model that then runs on a laptop. That contrast comes from Block and has not been verified, but it is the most concrete technical detail on offer. The Nvidia open-weight article covers what serving a model locally costs in practice, including licence and hardware caveats.
One correction to the launch coverage: the line “This is a GPT3 moment for speech” was posted by Thurairatnam, Block’s Dessa lead, on X. He is a maker, not an outside researcher, and it belongs in the same category as the page’s other superlatives. Those include “the most authentic full-duplex speech model ever built”, a hedged “as far as we know, Dessa is the world’s first AI you can have a realistic live conversation with that was trained only from human speech”, and “emotion is built into the model” as TechCrunch’s paraphrase of Block. None has been benchmarked, and the coverage contains no measured comparison between Dessa and any other voice model.
Why a payments company is doing this at all
Block’s stated reasons for Dessa are “curiosity (aka fun), better understanding between humans and AI, and safety”. It states no commercial plan, no roadmap and no Square integration.
The commercial context is real anyway. Square already sells AI-powered voice ordering, in which its system answers calls to a business number and routes orders to the point-of-sale, order manager and kitchen display. Square’s live help page also returned 403; that description comes from an archived capture. Nothing states that product uses Dessa.
TechCrunch’s closing analysis is that a payments company with its own on-device voice model “can power a lot more business and consumer experiences at a lower cost”. That is TechCrunch’s reasoning, not a Block plan, and the same caution applies to any wider claim about Square. What on-device inference reliably changes for a business of that scale is narrower: the per-interaction cost of running the model, the round-trip latency of a spoken order, and whether customer audio has to leave the device at all. Those are reasonable inferences from local inference in general. Block has not claimed them for Dessa, and no published artefact lets anyone test them.
What would change this article
Nothing in the release is dated, so there is no roadmap to watch. Dessa could be updated here if Block publishes a model card, a benchmark, a licence or a named external tester, or if it answers where the training conversations came from and on what terms.
