Wan3.0 Can Turn a PDF Into Video. The Draft Still Needs You

Alibaba's Wan3.0 has a headline-friendly number: it can generate a video up to 30 seconds long in one pass. The more interesting part is buried in the input list. Along with prompts, images, audio and existing clips, the model can read a PDF, slide deck, spreadsheet, text file or public web page and turn the material into video.[2]

That changes the job being automated. Most AI video demos start with a cinematic sentence and end with an attractive clip. Wan3.0 is also being pitched at the less glamorous work of converting material that already exists. Think product sheets, training documents, sales presentations and reports. The model is not waiting for someone to describe every shot from scratch; it can begin with the source document itself.

What actually launched

Alibaba Cloud's documentation describes Wan3.0 as one model for text-to-video, first-frame and first-to-last-frame animation, and reference-based generation. Output can run from two to 30 seconds when no source video is supplied, at 480p, 720p or 1080p. Audio is on by default, although the caller can disable it.[2]

The launch on August 24 widened access after a public beta that began earlier in the month.[4]

Coverage spread quickly across international and China-focused technology outlets, helped by the timing: Alibaba had just completed a $10.2 billion share offering earmarked for AI infrastructure.[5]

The useful distinction is between availability and openness. Wan3.0 is accessed through Alibaba Cloud services and its API, while the model weights remain closed. TNW notes that this breaks with the open releases that built much of the Wan family's developer following.[3] Anyone who wants to run the newest model on their own hardware, inspect it or fine-tune it independently does not have that option.

The document input is more than a file upload

The API accepts one file in formats including DOC, XLS, PPT, PDF, Markdown and plain text, with a maximum size of 100 MB. Many document types are limited to 50 pages. A public web page can be supplied instead, but it must be accessible without a login.[2]

From there, Wan3.0 is supposed to interpret the material and generate a video around it. The reference API also lets a request combine a written instruction with images, video or audio. Users can identify those assets by order inside the prompt, such as "Image 1" or "Video 1," to tell the model which object or character should do what.[2]

This is where the product starts to look less like a filmmaker and more like an automated production assistant. A ten-page presentation already contains a rough sequence, visual assets and the facts the creator wants to communicate. A spreadsheet has categories and comparisons. A product document has names, specifications and images. Treating those files as source material could remove a lot of copying between tools.

It could also create polished nonsense faster. A source file does not tell the model which facts deserve screen time, whether a chart needs caveats, or how much simplification is acceptable. The generated video still needs to be checked against the original. That review becomes especially important when the output includes numbers, instructions or a synthetic voiceover.

Thirty seconds matters, but not for the obvious reason

Wan3.0 doubles the 15-second ceiling reported for its predecessor.[3][5] Thirty seconds is long enough for a complete social clip, a short product explanation or one section of a training video. It also reduces the number of separate generations that an editor must stitch together.

Longer output creates a harder consistency problem. A five-second mistake can be discarded cheaply. In a 30-second shot, a drifting object, changed layout or mangled label may spoil much more material. Alibaba says Wan3.0 improves character detail, scene layout and motion across the clip, but independent benchmarking was not available at launch.[3] The demo reel is evidence of what the vendor selected, not a prediction of every user's result.

The API design also sets realistic expectations about speed. Video generation is asynchronous and typically takes one to five minutes, according to Alibaba's documentation. A client creates a task, polls for its status, then downloads a result from a link that expires after 24 hours.[2] This is a batch workflow, not instant video editing.

The price invites routine use

Published API rates are $0.05 per generated second at 480p, $0.10 at 720p and $0.20 at 1080p.[3] At those rates, a full 30-second clip costs $1.50, $3 or $6 before retries and editorial work. Those retries matter. The first result may be unusable, and a team testing several prompts can spend many times the sticker price on the clip it finally keeps.

Still, the pricing is low enough to make document-to-video tempting for ordinary business content. The likely competition is not only another frontier video model. It is the hour someone spends extracting bullet points, choosing stock footage, recording narration and adjusting a timeline. A rough generated first cut can have value even when it cannot be published unchanged.

That is also why the closed access matters. A company sending internal presentations or spreadsheets to a hosted model has to decide whether those files belong in that service. The API supports public URLs and encoded image data, but the broader data-handling decision sits with the customer. Sensitive documents should not become test prompts by accident.

What to test before trusting it

A sensible trial would use a short, non-confidential document whose facts are easy to verify. Keep the requested clip brief, ask for a simple visual sequence, and compare every claim in the output with the source. Test whether the model preserves names and numbers, not merely whether the motion looks good. Download successful results promptly because the API's result URLs expire after 24 hours.[2]

Wan3.0 does not make a PDF into a finished campaign with one click. It offers a potentially useful first draft of the video layer. The release is worth watching because it moves generative video closer to a common pile of work: the documents teams already have and rarely want to repackage by hand.

Sources

[2] https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-api-reference — Wan3.0 - Video Generation API Reference [3] https://thenextweb.com/news/alibaba-wan3-video-model-after-share-sale — Alibaba launches Wan3.0, its 30-second video model, days after raising $10bn [4] https://technode.com/2026/08/24/alibaba-launches-wan3-0-video-model-with-30-second-generation-and-document-input — Alibaba launches Wan3.0 video model with 30-second generation and document input [5] https://tech.yahoo.com/ai/articles/alibaba-wan3-0-ai-video-174537190.html — Alibaba Wan3.0 AI video model launch: 30-second video generation