Block says Dessa is a speech-to-speech model trained only on human conversation, running locally on an M4 Max or better. It is a research preview, and the page gives no parameter count.
Block says Dessa is speech-only, full-duplex, and local on an M4 Max or better. It is a research preview. The page gives no parameter count.
The short version
A speech-to-speech model trained from scratch on human conversation, with no text and no pretrained model, full-duplex, running locally on a MacBook Pro M4 Max or better. Block says that as far as it knows this is the first realistic live conversation model trained only from speech. Access is a research preview for safety researchers. The page states no parameter count, price, or public weights. A training rack named by The Decoder is not on the page this brief quotes.
The page is a preview, not a release
Block's page introduces Dessa as a speech-to-speech model. Audio in, audio out. It says the model was trained from scratch on human conversation, with no text and no pretrained models, and that as far as the company knows it is the first realistic live conversation model trained only from human speech. It says the model is full-duplex, so it listens and talks at the same time, and that inference runs faster than real time on a MacBook Pro with an M4 Max or better. It says a small group of AI safety researchers gets the research preview. The page does not give a parameter count, a price, a weight download, or a published evaluation. Those absences are the story as much as the adjectives.Nemotron's small model is a different release.
The Decoder adds a machine
The Decoder's October 24 story says the model was trained on a single NVIDIA GB200 NVL72. The company page fetched for this brief does not contain that sentence. Voice-industry writeups have repeated it. This page will not paste it into Block's mouth. If the company confirms the machine, the confirmation will be a company sentence. Until then the training computer is an outlet's report of a research preview, and the preview's own page is silent on it. A rack that is not on the page cannot become a specification.Phonon is an English voice with a published size.
The Verge listened to the demos
The Verge on October 23 describes clips Block published, including a conversation that stumbles, and it points at Jack Dorsey's company as the maker. TechCrunch's October 24 headline is the wire's other account of the same page. Neither article is a scored eval. A demo that sounds awkward is a demo that sounded awkward to that reporter. It is not a word-error rate, and it is not evidence the model fails a test the company never published. This page did not score the clips.Canva's planner is a different product.
What the company says is new
The claim that matters is the training restriction: speech only, no text, no pretrained model, and a live conversation rather than a pipeline that turns voice into text and back. Block says that pipeline throws away emotion, sarcasm, laughter, and pauses. That is an argument about product category. It is not a result on a shared test set. As far as we know, in the company's sentence, is a hedge, and this page keeps the hedge. A first is a first only inside the knowledge the speaker admits.Kimi's weights are a different model and a later date.
Who is in scope
Safety researchers Block says it is showing the preview to, and readers of the public page. People who want a local voice model they can download today are not holding a file. The M4 Max line is a hardware floor the company states for running it, not a store listing. A machine below that floor is outside the company's claim. A machine above it still needs the access the preview has not opened to the public.
What this page will not infer
It will not infer safety from the word safety. A preview offered to researchers is a distribution choice. It is not a finding that the model refuses a class of requests, and the page does not publish a refusal test. It will not infer a language list. It will not treat a full-duplex slogan as a measured overlap of speaking and listening. Asked directly, What does full-duplex mean on the page? Block says the model listens and talks at the same time, and that inference runs faster than real time. That is the company's description, not a latency measurement this page made.
Three dates, one page
The Verge is October 23. The Decoder and TechCrunch are October 24. The page is the object they are describing. Filing the newest headline as the day the model was trained would be a mistake. Filing the page as a general availability would be a larger one. The status word on the page is preview. Asked directly, Is Dessa generally available? Block describes a research preview aimed first at AI safety researchers. The page this brief uses does not list a price, an API, or a weight file.a context window is a different bound.
Why no size is a fact
Nemotron's brief, filed the same day, can say about 4 gigabytes because NVIDIA's page says so. This one cannot say a size Block did not print. The contrast is the point of linking them. Two voice-adjacent models, two kinds of disclosure. A reader who wants a number has a number on one page and a silence on the other. Asked directly, Does the company page name the training computer? No. The Decoder reports a single NVIDIA GB200 NVL72. That sentence stays The Decoder's until it appears on block.xyz.
The limit
No parameter count. No GB200 in Block's sentence. No benchmark invented from a demo. The record is the company's description of a speech-only preview and the hardware floor it names. The short answer this page will stand on is this: A speech-to-speech model trained from scratch on human conversation, with no text and no pretrained model, full-duplex, running locally on a MacBook Pro M4 Max or better. Block says that as far as it knows this is the first realistic live conversation model trained only from speech. Access is a research preview for safety researchers. The page states no parameter count, price, or public weights. A training rack named by The Decoder is not on the page this brief quotes.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
Is Dessa generally available?
Block describes a research preview aimed first at AI safety researchers. The page this brief uses does not list a price, an API, or a weight file.
Does the company page name the training computer?
No. The Decoder reports a single NVIDIA GB200 NVL72. That sentence stays The Decoder's until it appears on block.xyz.
What does full-duplex mean on the page?
Block says the model listens and talks at the same time, and that inference runs faster than real time. That is the company's description, not a latency measurement this page made.
Are The Verge's awkward clips a benchmark?
No. The Verge describes conversations Block published, including a clumsy turn. An anecdote about a demo is not a score.
Does Block claim a parameter count?
Not on the page this brief uses. The absence is part of the record. A newsroom that adds a size is adding something the company page did not.