Wan 3.0's open weights are still unconfirmed

Search for "Wan 3.0 open weights" and you will find confident answers in both directions. What you will not find is a checkpoint. As of 20 August 2026 there is no Hugging Face repository for Wan 3.0, no GitHub inference code, and no ComfyUI node. The reporting contradicts itself and the artifacts that would settle it do not exist.

The correct status is therefore: closed until Alibaba says otherwise. Not "probably open because the previous versions were." That prior looks reasonable and is, this quarter specifically, wrong.

What is actually confirmed

Wan 3.0 entered public beta on 6 August 2026 from Alibaba's Tongyi Lab. It generates native 30-second clips to 1080p with audio in a single pass. Its signature feature is Omni-Reference, which takes documents, spreadsheets, slides, PDFs and webpages as generation input. It is reachable through Model Studio and the Wan site. A full API is described as "soon," without a date.

That is the confirmed set. Everything about weights, licensing and self-hosting sits outside it.

The confusion is understandable, because the Wan family genuinely has an open track record. Earlier Wan releases are part of the reason "download a frontier-adjacent video model and run it yourself" became a realistic sentence in 2026 at all, alongside the other open releases this summer. Extrapolating that forward one version is the natural move. It is also exactly the move that just failed on the image side of the same company.

The four artifacts that would settle it

An open-weights release is not a vibe or an announcement. It is a set of files. Until these exist, treat the question as open:

1. A checkpoint, published under an organisation account the lab controls. Not a community re-upload, not a quantised fork of something adjacent, not a distillation someone trained to imitate the outputs.
2. A license file, naming actual terms. Apache-2.0 and a bespoke research license are both "open weights" in a headline and completely different in what they permit commercially.
3. Inference code, meaning a repository you can run the checkpoint with. Weights without code is a slower version of no weights.
4. A model card or technical report, describing architecture, training data posture and known limitations.

Downstream of those, a ComfyUI node is the usual signal that a release has been absorbed by the ecosystem rather than merely posted. None of the four exists for Wan 3.0 today.

Note the ordering trap in point two. "Open weights" answers can I download it. The license answers can my business use the output. Those are separate questions and only the second one pays. A checkpoint you can download under terms that exclude your jurisdiction, your use case or your revenue tier is not a self-host plan, and the license file is the only place that gets decided. Read it before you benchmark anything.

Alibaba already broke this pattern once this quarter

This is the part that should recalibrate the prior.

Qwen-Image-3.0 released on 21 July 2026 and went generally available on 5 August. Its instruction window jumped to 4,500 tokens, roughly 4.5 times Qwen-Image-2.0, aimed at dense text rendering: newspaper pages, multi-panel infographics, legible maths.

It shipped with no weights, no license, no technical report and no model card.

That is a hard break from the pattern. Qwen-Image 1.0 and 2.0 were Apache-2.0 releases with same-day technical reports. Anyone whose stack depended on self-hosting a Qwen image model discovered at 3.0 that the path had closed, in the same quarter that Wan 3.0's status became contested.

One company, two model lines, one confirmed reversal. "Alibaba open-sources its models" was true and is now, at minimum, not reliable. If you were planning a self-host migration on the strength of that pattern, the evidence base for it got materially weaker four weeks ago.

Writing a plan that survives the rumour being false

The goal is not to predict correctly. It is to make being wrong cheap. Five rules do most of the work.

1. Make hosted the default and self-host the branch. Write the plan as "we run hosted, and if these four artifacts appear under acceptable terms, we evaluate migrating." Not "we self-host, and hosted is the stopgap." The two read similarly and behave completely differently under uncertainty, because only one of them leaves you with a working pipeline when the release does not happen.

2. Give every "when weights drop" line a date and an owner. Any plan step conditional on an unconfirmed event needs an explicit falsification date. If no checkpoint exists by 1 November, we stop tracking this and re-plan. Without that date the plan does not fail, it just quietly persists, and six months later someone is still waiting.

3. Buy no hardware on an unconfirmed release. Rent for evaluation, always. A GPU purchase justified by a model that never ships is the single most expensive form of this mistake, and it is unrecoverable in a way that a wasted week of rental is not. The honest self-host math is unforgiving even when the weights do exist.

4. Specify the capability, not the model. Write the requirement down as "30-second single-pass generation with committed audio at 1080p" rather than "Wan 3.0." Stated as a capability, it has multiple candidate paths — including models already callable, and including a planned two-shot sequence stitched at a cut the edit wanted anyway. Stated as a model name, it has exactly one path and no fallback.

5. Check the license before the benchmark, not after. Reversing this order is how teams burn a week of evaluation on a checkpoint they were never permitted to deploy. The license takes ten minutes to read and eliminates candidates faster than any test.

What to run in the meantime

The callable end of the family is Wan 2.7, which generates from text with strong motion consistency, optional custom audio input and prompt rewriting. If your interest in Omni-Reference was really about conditioning generations on assets you already own, Wan 2.7 reference-to-video is the closer callable analogue and worth testing against the actual brief. The rest of the family sits alongside both.

That is the useful reframing. Most people wanting Wan 3.0 want one of two things: longer single-pass shots, or richer reference conditioning. Both have callable approximations today. Neither requires a checkpoint that may not exist.

And if the weights do land under workable terms, the migration is a research project you start then, from a working pipeline, with real output to compare against. That is a much better position than having spent three months waiting.

FAQ

Why do sources disagree about Wan 3.0's weights?

Because the earlier Wan releases were open and a lot of coverage extrapolated forward, sometimes phrasing an expectation as a fact. Once one outlet states it, aggregators repeat it. The tell is that no source links a checkpoint, which is the one thing an actual release would make trivial to link.

Does the Qwen-Image-3.0 closure mean Alibaba has stopped open-sourcing?

It means one model line closed at its third major version after two open ones. That is a data point about a policy that is clearly not uniform across the company, not a confirmed company-wide policy. It is enough to remove "Alibaba always opens its weights" from your list of safe assumptions.

If weights appear, what should I check before planning a migration?

License terms first, including any territorial or commercial restrictions. Then the actual hardware requirement at your target resolution and length, measured on rented hardware rather than taken from the announcement. Then whether a LoRA or reference-conditioning approach on a hosted model already gets you close enough that the migration is not worth the operational cost.

Is there any read on how likely the open release is?

Not one worth acting on. Alibaba has shipped open video weights before and shipped a closed image model this quarter. Both precedents are real and they point in opposite directions. Anyone offering you a probability here is guessing, which is precisely why the plan should be structured so the answer does not matter much.