Alibaba launched Wan3.0, its latest video generation model, on Monday, August twenty fourth, 2026, one day after pricing a placement of seven hundred and ten million new shares at one hundred and twelve Hong Kong dollars and seventy cents, raising eighty billion Hong Kong dollars, or roughly ten billion US dollars. Reuters describes it as the largest primary follow on offering ever by a Hong Kong listed company. The model produces thirty second videos from documents, spreadsheets, slides and web pages. The detail that matters if you build on open weights is that Wan3.0 is not one of them.
The short answer
Alibaba launched Wan3.0 on August 24, 2026, a video model that generates thirty second clips from documents, spreadsheets, slides and web pages. It arrived one day after the company priced 710 million new shares at HK$112.70, raising HK$80 billion, which Reuters calls the largest primary follow on offering ever by a Hong Kong listed company. Alibaba says all net proceeds fund its full stack AI capabilities. Wan3.0 itself is API only: no weights have been published.
There are two ways to read a model launch that lands the morning after a ten billion dollar raise. One is that the company is confident. The other is that the model is what the raise has to justify. Both are probably true here, and neither is the detail most likely to affect your work.
What Wan3.0 does
Alibaba Cloud says Wan3.0 generates thirty second videos, and that it accepts documents, spreadsheets, slides and web pages as input. That second half is the more interesting claim, because it changes the product shape.
Prompt to video is a creative tool. Document to video is a conversion tool, and it aims at a different buyer entirely: the person with a slide deck who needs a promotional clip, the team with a product page that wants a demo reel. Alibaba lists the beta use cases accordingly, naming short drama and film production, advertising and marketing, tourism promotion and music video creation. The model has been in public beta since August 6, 2026, on Alibaba Cloud Model Studio.
Thirty seconds is also worth noting on its own. Length has been the binding constraint on generated video for most of the current generation of models, where clips are measured in single digit seconds and anything longer is stitched. If the thirty second figure holds up as a single coherent generation rather than an assembly, that is a real capability step.
The part that changes your options
Here is the thing that did not appear in most of the coverage. Wan3.0 is an API product. The weights have not been released.
We checked the two places they would be. The Wan-AI organisation on Hugging Face lists twenty seven model repositories, and the highest version number present anywhere in them is 2.2. There is no Wan3.0 repository. Alibaba's last open weight video flagship remains Wan 2.2, published under Apache 2.0 in 2025 and still perfectly usable, but a generation behind what was launched today.
That distinction is easy to lose in a headline and expensive to discover late. If your reason for caring about a Chinese video model was that the previous ones came with downloadable weights and a permissive licence, this release does not extend that. It is available to call, not to run.
Nobody has said this is a policy change. Alibaba is one of the most prolific publishers of open weight models in the industry, across both the Qwen language family and the Wan video family, and none of that has been withdrawn. But the observable state today is a closed flagship and an open predecessor, which is a pattern worth watching rather than a conclusion worth drawing. Compare it with labs still shipping their frontier work openly, such as Moonshot's Kimi K3 at 2.8 trillion parameters.
What the ten billion dollars is for
The financing is not incidental context, it is the same story. On August 23, 2026, Alibaba priced a placement of seven hundred and ten million newly issued ordinary shares at one hundred and twelve Hong Kong dollars and seventy cents, raising eighty billion Hong Kong dollars, roughly ten billion US dollars. Reuters reports it as the largest primary follow on offering ever by a Hong Kong listed company.
Alibaba states that one hundred per cent of the net proceeds go to its full stack AI capabilities, including expanding and enhancing its AI infrastructure. The timing is explained by the quarter that preceded it: earnings fell roughly seventy five per cent year over year, driven largely by AI capital expenditure. A company spending at that rate either raises or slows down, and this is the choice it made.
That context is also the most plausible explanation for a closed flagship. Training and serving a video model at this scale is expensive in exactly the way the placement is meant to fund, and an API is a revenue line where a weight release is not.
The practical read
If you want the output, it is there. The API is live, thirty second generation from a document is a capability nothing open currently matches, and for a marketing or content pipeline that may be the whole decision.
If you need to run the model yourself, for cost, for data handling, for offline work or for fine tuning, then today's launch does not apply to you and Wan 2.2 is still your ceiling. The mistake to avoid is building against the API on the assumption that weights will follow, because no such commitment has been made. Plan for what is published, not for what would be convenient.
Sources and further reading
- Alibaba launches Wan3.0 AI video model after $10 billion share sale, Reuters via Yahoo Finance, August 24, 2026
- Alibaba Group Holding Ltd, Form 6-K, placement pricing announcement, SEC, August 23, 2026
- Wan-AI model repositories, Hugging Face
Frequently asked questions
Can I download the Wan3.0 weights?
Not as of August 24, 2026. The Wan-AI organisation on Hugging Face lists twenty seven model repositories and the highest version number present is 2.2, with no 3.0 repository. Wan3.0 has been available through Alibaba Cloud Model Studio since its public beta on August 6, which makes it an API product rather than a released model. Alibaba has not said whether weights will follow. If your plan depends on running the model yourself, the newest thing you can actually run is Wan 2.2, which is published under Apache 2.0.
What does Wan3.0 actually do that is new?
The headline capability is length and input breadth together. Alibaba Cloud says the model generates thirty second videos, and that it takes documents, spreadsheets, slides and web pages as inputs rather than only a text prompt or a starting image. Feeding it a document and getting a video out is a different product shape from prompt to video, and it points squarely at marketing, training material and short form promotional content. Alibaba says it has been used in short drama and film production, advertising and marketing, tourism promotion and music video creation since the beta opened.
How big is the share placement and what is it for?
Alibaba priced seven hundred and ten million newly issued ordinary shares at one hundred and twelve Hong Kong dollars and seventy cents each on August 23, 2026, raising eighty billion Hong Kong dollars, about ten billion US dollars. The company states that one hundred per cent of the net proceeds go to its full stack AI capabilities, including expanding and enhancing its AI infrastructure. Reuters reports it as the largest ever primary follow on offering by a Hong Kong listed company. It follows a quarter in which Alibaba's earnings fell roughly seventy five per cent year over year, largely on AI capital expenditure.
Does a closed Wan3.0 mean Alibaba is done with open weights?
There is no statement to that effect, and reading a policy change into one release would be premature. Alibaba has been among the most prolific publishers of open weight models, across both the Qwen language family and the Wan video family, and those repositories are still there. What can be observed is a specific gap: the current flagship video model is API only while the last open flagship is Wan 2.2 from 2025. Whether that gap closes later, stays open, or becomes the pattern is not something anyone can answer from the announcement.
What should I do if I was planning to build on Wan3.0?
Decide first whether you need the weights or only the output. If you need the output, the API is available now through Alibaba Cloud Model Studio and the thirty second length is a genuine capability difference. If you need to run it yourself, for cost control, data residency, offline use or fine tuning, then Wan3.0 is not currently an option and Wan 2.2 under Apache 2.0 is your ceiling. Building a pipeline against an API on the assumption that weights will follow is the mistake to avoid, because nothing has been promised.