Alibaba's Wan3.0: your spec sheet is now video input
Alibaba shipped Wan3.0, a video model that takes PDFs, slides, and spreadsheets as input and returns 30-second clips. Product marketing just got a new source file.
Most video models want a prompt. Alibaba's Wan3.0 wants your documents. Alibaba rolled the model out of beta on Monday, Reuters reported, and the input list is the interesting part: documents, spreadsheets, slides, and web pages go in, and up to 30 seconds of video comes out. For anyone maintaining a catalog, that changes which asset is the source of truth.
What actually happened
Wan3.0 went generally available on August 24 after a public beta earlier this month. Maximum clip length is 30 seconds, double the 15-second ceiling of its predecessor Wan2.7. Per TechNode, it accepts text, images, video, audio, web pages, PDFs, and PowerPoint files, and it can carry characters, props, spatial layout, and style across from reference material. It also handles multilingual voice output and — unusually for a video model — stable rendering of UI elements and motion graphics.
The launch came a day after Alibaba priced a $10 billion share placement to fund AI spending, which Reuters described as the largest primary follow-on offering by a Hong Kong–listed company. Alibaba is buying compute and shipping models on the same news cycle. We are not repeating the per-second pricing floating around in secondary coverage; we could not confirm it against Alibaba Cloud's own pricing, so treat cost as something to measure yourself in Model Studio.
Why document-to-video matters for your marketing pipeline
Your spec sheet becomes a reusable asset. The reason product video never scales for a small catalog is that every clip is a fresh creative project. When a PDF or a slide is a valid input, the marginal cost of the 200th SKU video collapses toward the cost of having accurate product data. That reframes the work: the leverage is in clean structured data, not in a video editor.
UI and motion graphics rendering is the sleeper feature. Text and interface elements are where generative video usually falls apart into garbled pixels. If Wan3.0 holds up here, software demo clips, feature explainers, and how-to segments become generatable instead of screen-recorded. Test that specifically before you believe it.
A Chinese-hosted model is a data decision. Running your catalog, pricing sheets, and unreleased decks through Alibaba Cloud is a data residency and disclosure question, not just a quality one. Decide what class of document is allowed through before someone in marketing uploads the roadmap deck. Wan's earlier weights have been openly released, so a self-hosted path may exist — check the license for the version you actually want.
Key takeaways
- Wan3.0 launched August 24, 2026 after a public beta earlier in the month
- Accepts PDFs, slides, spreadsheets, web pages, images, audio, and video as input
- Max clip length is 30 seconds, up from 15 in Wan2.7
- Claims stable UI and motion-graphics rendering — verify on your own screens first
- Alibaba priced a $10B share placement the day before to fund AI spending
- Set a policy on which documents may be uploaded before marketing decides for you
Video at catalog scale is a data problem wearing a creative costume. We build the product data layer that makes generated assets accurate instead of embarrassing. See how we structure catalogs, or bring us your SKU sprawl.
- #video-generation
- #alibaba
- #product-marketing
- #content-automation
- #open-models
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Instinct's AI assistant asks for a perpetual license
Instinct's AI assistant wants your inbox, calendar, screen, and keystrokes — plus a perpetual license to it all. A checklist for vetting any agent you connect to work accounts.
Read itHugging Face weighs a $13B sale: mirror your weights
Hugging Face is reportedly exploring a sale at $13B or more. If your build pulls models and datasets from the Hub at deploy time, that's a single point of failure.
Read it