Developer

Dessa AI Open Source Alternatives Developers Can Run Today

2026-11-12 · 7 min read · By DessaAI Daily

The searches keep growing — Dessa on GitHub, Dessa voice GitHub, Block Dessa GitHub — and they all land on the same fact: there is no open-source Dessa to download. Block has released no code, no weights and no licence. If you are a developer who wants to run a full-duplex voice model on your own machine this week, this guide answers the follow-up question honestly: what can you actually run today, how does it differ from Dessa, and what should you check before you build on it?

Why there is no Dessa repo to clone

In our guide to Dessa AI on GitHub we set out the status in full, and it has not changed. Dessa, announced on October 23, 2026, is a research preview from a small Block lab. Block has published demo recordings, a waitlist and a route for a small group of AI safety researchers to request early access. It has published no repository, no weights and no open-source licence — a position we also checked from the model-hub side in Dessa AI on Hugging Face and from the weights side in Dessa AI open weights.

That gating is a deliberate choice, and Block's own framing explains it: a voice that can hold a live, emotional, interruptible conversation raises impersonation and scam risks, so outside safety researchers go first. Whether you agree with that order or not, the practical consequence is simple. Anything on a code or model site today that claims to be Block's Dessa weights or Block's official Dessa repo should be treated as unofficial until Block itself announces a release.

The open counterpart: Moshi

The clearest open alternative is Moshi, from the Paris research lab Kyutai. Moshi is a realtime, full-duplex spoken dialogue system, and Kyutai released its weights openly — which means a developer can download it, run it, inspect it and publish results that other people can check. That inspectability is the opposite experience from a gated preview, and for many research and prototyping tasks it is the decisive advantage.

The foundation, however, is different in a way that matters. As we explained in Dessa AI vs Moshi, Moshi is built on Kyutai's Helium text language model and then extended into audio. Text is in its foundation. Dessa's claim is more radical: trained from scratch on human conversation audio only, with no text and no pretrained models, learning language the way Block argues babies do — by listening. Running Moshi therefore tests full-duplex conversation today; it does not test Block's speech-only hypothesis, and it should not be presented to users or managers as a stand-in for Dessa itself.

NVIDIA's PersonaPlex, described as a Moshi fine-tune, shows the other benefit of an open release: other teams can build variants with different personas and publish them for scrutiny. We compared that route in Dessa AI vs NVIDIA PersonaPlex. None of that ecosystem exists around Dessa yet, because there is nothing public to build on.

How to choose what to run

Choose by the question you need to answer. If you need to learn how full-duplex timing, interruption and backchanneling feel in a real system, an open model you can run today is the only way to learn it hands-on — a demo recording cannot teach your hands what latency and overlap do to a product. If you need a supported product for customers today, an open research system may not be enough either, and the commercial options in Dessa AI alternatives you can use today are the fair comparison. If you specifically need to evaluate Dessa — its speech-only training, its emotion, its local inference claim — then no alternative answers that question, and the honest plan is the waitlist and the safety researchers' published findings, when they appear.

Whatever you run, check the ordinary open-source basics before you invest: the actual licence in the repo, the hardware the model realistically needs, whether the project is maintained, and whether the download in front of you comes from the organisation that actually released it. Those checks matter more, not less, when a famous closed model's name is being used by impostor uploads trading on its announcement.

The local-hardware point, kept honest

Block says Dessa runs faster than real time locally on a MacBook Pro with an M4 Max chip or better, after training on a single NVIDIA GB200 NVL72. That is a claim about a model you cannot yet obtain. Open models invert the trade: the weights are obtainable, and the hardware question becomes one you can measure yourself, on your own machine, with your own prompts. For a developer, that measurability is worth a great deal — it turns a launch-page claim into a benchmark you ran.

Block has published no parameter count, no benchmarks and no training hours for Dessa, so no honest article can rank it against a runnable open model on performance. What can be ranked today is access: open weights you can run, versus a preview you can only listen to. If Block later releases code or weights under a clear licence, we will update our GitHub and open-weights guides the same day and link the official repo from here. Until then, run what is actually open, name it accurately, and do not call it Dessa.

Sources: Block's official Dessa page (block.xyz/dessa) for Dessa's status, training and local-inference claims; Kyutai's Moshi paper and project materials (arXiv:2410.00037) for Moshi's open release and Helium foundation; DessaAI Daily's GitHub, Hugging Face and open-weights guides, checked November 2026.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗