Developer

Dessa AI Open Weights: Has Block Released the Model Weights?

2026-11-02 · 7 min read · By DessaAI Daily

Someone searched Google for "dessa open weights" this week — and it is exactly the right question to ask about Block's speech-only model. So here is the straight answer first: Block has not released Dessa's weights, and it has not announced any plan to do so. There is no download, no licence, no model card and no repository attached to the October 23, 2026 announcement. Dessa is a research preview you can listen to in recorded demos and join a waitlist for — not a model you can obtain and run yourself today.

That matters because "open weights" has become the dividing line in voice AI. Some rivals publish their weights for anyone to download and build on; others keep everything behind a hosted service. Dessa, for now, sits in neither camp in public: Block has described how it was built in unusual detail, while keeping the model itself entirely in-house. This article separates what Block has actually said about openness from what people are hoping it will do next.

What Block has — and has not — announced about the weights

Start with the verified list. In its announcement, Block said Dessa was trained from scratch on human conversation audio only, with no text and no pretrained models, on a single NVIDIA GB200 NVL72 system, and that inference runs faster than real time locally on a MacBook Pro with an M4 Max chip or better. The team is a small Block lab led by Ragavan Thurairatnam, reporting to Jack Dorsey, and early access is going first to a small group of AI safety researchers, with a public waitlist at block.xyz/dessa.

Now the other list, which is just as important. Block has not announced: model weights, an open-source or open-weights licence, a parameter count, training-data hours, benchmarks, pricing, an API, or a public release date. Our guide to the Dessa AI API covers the developer side of that same gap: as with the API, anyone claiming to offer a Dessa download today is offering something Block did not release. The current status is recorded in one place on our Dessa Tracker.

Why the local-running detail fuels the open-weights question

It is easy to see why searchers connect the dots. A model that runs locally on a high-end laptop sounds like a model you might one day download — and Block itself highlighted that local, faster-than-real-time inference rather than a cloud-only story. Our explainer on why Dessa runs on a laptop walks through what that claim does and does not imply. Running locally inside Block's own testing, however, is not the same as releasing weights publicly. Plenty of research models run on local hardware inside a lab for months before any decision is made about distribution — if a public distribution decision happens at all.

There is also a safety reason to expect caution, and Block has signalled it plainly: the first outside users are safety researchers, not developers or hobbyists. A voice model that hears and produces emotion, sarcasm and laughter — and that can sound convincingly human in a live conversation — raises impersonation and fraud questions that a text model does not raise in the same way. Releasing weights would put that capability beyond recall. Whether or not that is Block's reasoning, the announced order of access points away from an imminent open release.

How Dessa compares with a genuinely open-weights rival

The clearest contrast is Kyutai's Moshi, the open-weights full-duplex voice model built on a text language model. Moshi's weights are public, so researchers can inspect, fine-tune and self-host it — and they can also verify its limits for themselves. Our comparison, Dessa AI vs Moshi, lays out the trade-off: Moshi gives the world the artefact and the openness that comes with it, while Dessa's speech-only training approach remains, for now, a described method backed by selected demo recordings rather than an independently testable model.

That distinction should shape how you read every Dessa claim. Full-duplex conversation, backchanneling with "yeah" and "got it," laughing mid-sentence, adjusting volume when asked — all of that comes from Block's own live, unsteered demo recordings. They are impressive recordings. They are not the same kind of evidence as weights that outside researchers have run, measured and tried to break. Until Block either widens hands-on access or publishes more technical detail, honest coverage has to keep those two categories separate, and so does this site.

What to watch for if the position changes

If Block moves toward any form of release, the signals will be specific and checkable, not vibes. Watch for: a named licence; a parameter count and model card; published benchmarks others can reproduce; an API with documentation and terms; or a widening of early access beyond safety researchers. Any one of those would be a genuine change to the status board — and the day one appears, it will be headline news here, with the source linked.

Until then, treat the answer as settled: no Dessa open weights exist in public, no date for them has been announced, and no third-party site can give you what Block has not released. If you are new to the model itself, start with What Is Dessa AI? for the full picture, and join the official waitlist if you want to be in line when access — in whatever form Block chooses — actually opens.

Sources: Block's official Dessa page (block.xyz/dessa), October 23, 2026; verified-facts list in this site's editorial bible. Nothing about weights, licensing or release timing beyond the above has been announced by Block.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗