Explainer
Dessa AI Parameter Count: How Big Is Block's Speech Model?
How big is Dessa? It is a fair question for anyone who found Block's speech model through a model hub, a GitHub search or a Hugging Face search. The honest answer is short: Block has not announced a Dessa AI parameter count, a model size or an architecture breakdown. No model card has been published. This article explains what is actually known, why the one real clue — that Dessa runs locally on a high-end laptop — does not turn into a number, and how to spot invented sizes.
The record: no parameter count has been published
Dessa was announced by Block on October 23, 2026 as a full-duplex, speech-to-speech model trained from scratch on human conversation audio only — no text and no pretrained models. Block's official page describes what the model does: it hears emotion, sarcasm and laughter, it backchannels, it knows when to interrupt, and inference runs faster than real time. What the page does not give is the kind of specification sheet developers are used to from open model releases. There is no parameter count, no layer count, no context figure presented as a verified Dessa specification, and no downloadable file whose size you could inspect.
That absence matters because parameter counts are normally published by the model's owner, in a paper, a model card or a repository. None of those artifacts exists publicly for Dessa today. Our article on Dessa AI on Hugging Face and GitHub covers the repo question directly: there is no official model page you can open to read a size off. Likewise, our Dessa AI benchmarks explainer notes that Block has published no benchmark table either. Size and scores are both in the not-announced column.
What the M4 Max clue does — and does not — tell you
The closest thing to a size clue is hardware. Block says Dessa runs locally on a MacBook Pro with an M4 Max chip or better. That tells you the model, as Block demonstrated it, can run on a powerful consumer laptop rather than only in a data centre. It does not tell you how many parameters it has. A hardware minimum is not a model size: memory, speed, optimisation and how a system is packaged all affect what hardware a model needs, and Block has not published the breakdown that would let anyone convert 'runs on an M4 Max' into a trustworthy parameter figure.
Our guide to the machine that runs Dessa and our explainer on why Dessa runs on a laptop cover that local-running claim in full. Both start from the same limit: local running is a verified Block claim, model size is not a verified Block disclosure. Treating the first as proof of the second is guessing, however confident the guess sounds.
Why speech-only makes size comparisons tricky anyway
Even if Block published a number tomorrow, comparing it directly with a text large language model would need care. Dessa did not learn language from text. Block says it learned from human conversation audio, from hundreds of times fewer words than LLMs typically use, trained on a single NVIDIA GB200 NVL72 system. A speech-only model's parameters are doing a different job from a text model's parameters: handling sound, timing, emotion and turn-taking directly, rather than predicting the next written token. Our how Dessa AI works explainer walks through that difference.
That is also why third-party listings that assign Dessa a neat size, a token limit or a benchmark score should be read sceptically. If the figure does not trace to Block's official page or a Block-published artifact, it is not a Dessa specification — it is a database filling a blank. Block has left that blank blank.
How to check any Dessa size claim you see
Use three checks. First, ask where the number was published: is there a Block page, paper or official repository behind it? Second, ask whether weights or a model card exist to support it — as our open weights article records, Block has not released weights. Third, check the date: Dessa is still a research preview, with early access for a small group of AI safety researchers first and a waitlist at block.xyz/dessa. A model in that stage can absolutely have its size announced later. Until Block does, the correct entry for parameter count is the same as for price, API and release date: not announced.
Sources: Block's official Dessa page at block.xyz/dessa, checked November 14, 2026; DessaAI Daily's verified-facts record. No parameter count, model size, pricing, API, weights or release date had been announced as of that check.
Frequently asked questions
What is Dessa AI's parameter count?
Block has not announced it. No parameter count, model size, architecture breakdown or model card has been published for Dessa, so any number circulating online is not a verified Block figure.
Is Dessa listed on Hugging Face with its size?
No official Dessa model page with a size or weights was found in our checks. The Hugging Face question is covered in our dedicated article on whether there is an official model page or repo.
Does running on an M4 Max prove Dessa is small?
No. Block says Dessa runs locally on a MacBook Pro with an M4 Max or better and runs faster than real time. That is a hardware floor, not a parameter count, and it does not reveal the model's size.
🚀 Get Dessa launch alerts
Dessa is still a research preview. Join the free DessaAI Daily list and we will email you the day access opens — plus the week's biggest Dessa stories, once a week. No spam, no hype.
Form not loading? Subscribe on the signup page →