Wan 3.0 is Alibaba Tongyi Lab's video generation and editing model, built as an all-in-one where text, images, audio, video and even office documents can feed a generation. Its public beta opened on August 6, 2026 across the group's own services. One pass renders up to 30 seconds of footage, where Wan 2.7 topped out between 2 and 15 seconds. API billing runs per second of output, in yuan.
Wan 3.0 merges into a single model what Wan 2.7 split across four specialists: text-to-video, image-to-video, reference-based generation and editing. The 30-second shot comes out in one go, with no clips stitched afterwards, which makes continuous one-take camera moves possible.
Two helpers round out the package. The first suggests a clip length matched to your prompt (no more picking a duration blind), the second extends a video you have already rendered. Alibaba also highlights stable characters, props and sets from the first frame to the last.
Wan 3.0 accepts doc, xls, ppt, pdf and md files, plus web pages, as creative references. You drop in a quarterly report, describe the tone you want, and the model derives a clip from it without a summary-then-storyboard detour. Few AI video generation tools go beyond text and image as inputs.
| Accepted input | What Wan 3.0 does with it |
|---|---|
| Text | Describes the scene, tone and camera moves |
| Image | Acts as a starting point or style reference |
| Audio and video | Guide the pacing and motion of the render |
| Document (doc, xls, ppt, pdf, md) | Becomes the script of the clip |
| Web page | Supplies the content to put on screen |
Wan 3.0's public beta is rolled out across several Alibaba services rather than one dedicated app, and access usually goes through a China-region account. The API, gradually opening to developers, bills per second of output: ¥0.3 at 480p, ¥0.6 at 720p and ¥1.2 at 1080p, roughly $0.04 to $0.17 per second. A 30-second 1080p clip lands around ¥36, close to $5. Those figures date from the beta, so Alibaba's published rate card is the one to trust.
On open weights, the company has committed to an Apache 2.0 release, yet nothing is online so far: the line's last published weights stop at Wan 2.2, and the community wants proof after the Wan 2.5 episode, announced as open and never shipped. Alibaba also concedes two areas left to improve, audio quality and the accuracy of on-screen text.
Wan 3.0 has no standard subscription: the public beta runs through Alibaba's services and the API bills per second of generated video, at a rate depending on resolution. Since no open weights are available yet, there is no version you can run for free on your own hardware for now.
Alibaba has pointed to an Apache 2.0 release, but no Wan 3.0 weights are online. The series' last open weights date back to Wan 2.2, and the Wan 2.5 precedent, presented as open then never delivered, has the community waiting for a verifiable Hugging Face repository before believing it.
On shot length, Wan 3.0 takes the lead with 30 seconds generated in one block. It had not yet appeared in ranking arenas when its beta opened, though, and its audio remains a weak spot Alibaba openly admits. Its document-based input, on the other hand, has no direct equivalent among rivals.
Access to the beta generally assumes a China-region account, billed in yuan, since the trial channels are Alibaba's Chinese services. The developer API is opening up in stages, and third-party platforms already hosting the Wan series have said they plan to add it to their catalogs.
Verdict: Ad and social teams churning out campaign variants get a model that can absorb a full brief, PDF included, and hand back one continuous shot; for a polished soundtrack, a dedicated audio pass fills the gap while the announced fixes arrive.
