Explainer
Dessa AI on GitHub: Is Block's Voice Model Open Source?
For developers, the reflex is automatic: hear about a new model, search GitHub for the repo. Dessa, Block's speech-only voice model announced on October 23, 2026, sends a lot of people down that path — so here is the verified state of play. Block has published no Dessa code, no GitHub repository and no open-source license. This article explains what that means, how code differs from the weights and API developers also ask about, and what you will actually find when you search.
There is no official Dessa GitHub repository
Block's official Dessa page — re-read today — links to its demos, its research description and its safety-researcher access route. It does not link to a code repository, because none has been released. There is no official Dessa organisation publishing the model's code, no reference implementation, no inference script and no license file to read. That places Dessa at the closed end of the spectrum we described in what Dessa AI is: a research preview whose workings are described in prose and demonstrated in recordings, but not inspectable in code.
This matters practically. Without a repo, there is nothing to clone, fork, audit or run. Anyone who needs to evaluate how Dessa is built has only Block's claims: trained from scratch on human conversation audio only, no text, no pretrained models, on a single NVIDIA GB200 NVL72, running faster than real time locally on a MacBook Pro with an M4 Max chip or better. Those claims may well hold up — but code is how the field usually checks, and Block has not offered it.
Code, weights and API: three different doors, all closed
Developer searches often blur three separate things, so it is worth separating them. Code is the training and inference implementation — not released. Weights are the trained model files you could load and run — not released either, as our open weights explainer covers in detail. An API is hosted access without any files at all — also not announced, as our API explainer sets out. Block has opened none of the three doors. A GitHub search cannot find what does not exist, and the absence is a policy choice, not a search failure.
What you will find on GitHub instead
Search anyway and you will find things — just not Block's model. Expect third-party projects using the name: demo write-ups, link collections, speculative wrappers, and unrelated software that shares the word "Dessa" (the name predates the model; the original Toronto startup and other projects used it too). None of that is Block code, and a repository calling itself an "implementation" of Dessa cannot be one in any verifiable sense, because the architecture details, training data and weights it would need have never been published. Treat downloads from such repos with the same caution as any unofficial binary: at best they are fan projects, at worst they are borrowing a famous name. The genuine open point of comparison is Kyutai's Moshi, whose weights were released openly — the contrast at the heart of our Dessa vs Moshi comparison.
Why Block is keeping the code closed — for now
Block has stated its sequencing plainly: a small group of AI safety researchers gets early access first, to find risks the team has not anticipated, before any wider release. A human-sounding, full-duplex voice raises impersonation and scam questions that a payments company — Block owns Square and Cash App — has concrete reasons to take seriously. The lab's own history points the same way: it descends from the Toronto Dessa team whose 2019 Joe Rogan deepfake demonstrated voice-cloning risk early, a story we tell in the history of Dessa. Whether code or weights follow the safety phase is genuinely unknown; Block has announced neither a release nor a refusal, only a research preview.
What developers can do today
If you want to build with full-duplex speech now, work with what is actually published: Moshi's open weights for research, or the hosted voice APIs covered in our alternatives coverage, which ship documentation and pricing today. If you want to follow Dessa itself, the official page and the waitlist are the only primary sources, and this site tracks every change to the release picture — code, weights, API or price — on the day Block announces it. A Dessa GitHub link, if one ever exists, will appear there first. Until then, the most accurate README is the one Block has effectively written: not released.
Sources: Block's official Dessa page (block.xyz/dessa), re-read November 6, 2026 — no code repository, weights, API or license published; Kyutai's Moshi paper and project materials (arXiv:2410.00037) for the open-weights contrast; The Verge and TechCrunch, October 2026.
Frequently asked questions
Is Dessa AI on GitHub?
No official repository exists. Block has published no Dessa code, no GitHub repo and no open-source license. Repositories using the Dessa name are third-party projects, not Block's model.
Is Dessa AI open source?
No. Block has released neither source code nor model weights, and has announced no license. Dessa is a closed research preview; early access goes first to a small group of AI safety researchers.
Where can I find full-duplex voice code I can actually use?
Kyutai's Moshi was released with open weights and is the closest open comparison to Dessa. Hosted voice APIs from established providers are the other buildable-today route; Dessa itself has no API yet.
🚀 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 →