TL;DR
- Formal Release: Alibaba released the 30-second Wan3.0 AI video model on August 24 after its beta; the Model Studio API remains in preview.
- Source Inputs: One request can use text, images, video, audio, a document, or a public webpage, with generated audio enabled by default.
- Developer Access: Approved developers need Model Studio access and API credentials that match a supported region.
- Hosted Model: Developers could use Alibaba’s hosted Model Studio service, while its Wan repositories offered no model files for self-hosting by August 25.
- List Price: List rates are $0.05, $0.10, and $0.20 per output second for 480p, 720p, and 1080p video, respectively.
Alibaba Cloud has released Wan3.0, its hosted AI video model, on August 24, two weeks after an application-only public beta. The launch added several Chinese consumer and creator services, although their eligibility rules remained unspecified. Approved developers can also call the preview API through Alibaba Cloud’s Model Studio with region-matched credentials. Wan3.0 makes clips up to 30 seconds, twice the limit of Wan2.7’s text-to-video path, from prompts, media, documents, or public webpages.
Wan3.0 can turn a longer collection of source material into one longer video job without requiring the user to reduce a document or webpage to a text prompt first. Alibaba bills each generated second, so the doubled duration ceiling also doubles the list-price exposure at a given resolution.
Native 30-Second Video Generation
Generate videos up to 30 seconds in a single run with Wan3.0. pic.twitter.com/eQ1JnWvpzf— Wan (@Alibaba_Wan) August 6, 2026
Longer Clips From More Than Prompts
The Wan3.0 API supports text-to-video, first-frame and first-and-last-frame generation, and a reference mode that can draw on images, video, audio, a file, or a public web link. A request can include as many as 10 reference images, five videos, and five audio clips. The reference videos and audio can each total no more than 15 seconds.
Documents and webpages give the model a different kind of starting point. A user can attach one file, up to 100 MB, or one publicly accessible webpage that does not require a login. Supported files include PDF, Word, PowerPoint, Excel, plain text, Markdown, and Apple’s Keynote, Pages, and Numbers formats. Documents in the listed paged formats can contain up to 50 pages. The file and web-link options are mutually exclusive, making this a way to ground one generation job in a structured source rather than a general document library.
Output runs at 30 frames per second in 480p, 720p, or 1080p. Without an input video, the requested duration can range from two to 30 seconds. When a video is supplied, the source video’s length plus the generated continuation cannot exceed 30 seconds. Audio generation is on by default and can be switched off without changing the listed rate.
Alibaba says Wan3.0 also carries forward video-editing abilities introduced with Wan2.7, including changes to visuals, plot, and dialogue. The Wan2.7 text-to-video API provides the direct baseline: that path tops out at 15 seconds and lacks Wan3.0’s document-or-webpage input.
That’s just a glimpse of what Wan3.0 can do.
More demos below. 👇 pic.twitter.com/CsuQ8XbP7E— Wan (@Alibaba_Wan) August 6, 2026
The Formal Release Adds Services While the API Stays in Preview
Alibaba opened the public beta on August 6, with application access through Model Studio and Qwen Cloud. The formal release followed on August 24. Alibaba Cloud added China-facing access through services including the Qwen app, Qwen Creation for PCs, and IF Studio. Eligibility for those services remained unspecified.
For developers, Model Studio continues to provide a hosted, asynchronous API in preview. Access requires approval, and the API key, workspace, and endpoint must use the same region. Alibaba documents endpoints in Beijing, Singapore, and Virginia.
As of August 25, developers could use Wan3.0 through Alibaba’s hosted service, while the canonical Wan organizations on GitHub and Hugging Face offered no Wan3.0 model files for running it on their own systems. That differs from the prior lineage: Wan2.2 was released as an open-source update.
Price Rises With Resolution and Duration
Alibaba bills Wan3.0 by generated second rather than by request. At the international Model Studio list rates, a full 30-second output ranges from $1.50 at 480p to $6 at 1080p before taxes, discounts, or account-specific terms.
Wan3.0 international list rates
A customer’s final access and bill also depend on preview approval, account region, taxes, promotions, and any negotiated discounts. At a given resolution, using the full 30-second window costs twice as much as generating 15 seconds.
Wan3.0 Arrives as Alibaba Raises More Money for AI
Alibaba is selling Wan3.0 by output second while expanding the infrastructure behind its cloud AI business. The day before the model’s formal release, the company priced a new-share placement worth HK$80 billion, about $10.2 billion, with closing expected on August 26 subject to conditions. Alibaba said all net proceeds would fund full-stack AI, including infrastructure, research and models, and AI-native applications.
In the June quarter, AI Cloud and Compute Services revenue rose 45% from a year earlier to RMB48.437 billion, while capital spending rose 75% to RMB67.678 billion. The placement and financial report name no Wan3.0-specific allocation, so they explain the company’s investment setting rather than the model’s capabilities or performance.
A Transparent Test Finds Gains and Unwanted Cuts
Alibaba says Wan3.0 provides better visual fidelity, character consistency, facial detail, and physical behavior. The company identifies audio texture and the accuracy of on-screen text as areas that need further improvement, two practical limits for a model built around synchronized audiovisual output and document-based generation.
Model-hosting company Atlas Cloud says it ran 49 fixed-seed comparisons using the same prompts and starting frames across Wan3.0, Wan2.7, and Seedance 2.0. It reported stronger identity retention and physical behavior in several scenarios, alongside repeated unwanted cuts during shots that were supposed to remain continuous. Atlas Cloud sells access to the compared models, which gives its test a commercial interest.
Neutral leaderboards had not benchmarked Wan3.0 by the formal release. For creators who need one continuous 30-second shot, Atlas Cloud’s repeated unwanted cuts are the clearest present warning about how the longer output may behave in practice.
