Alibaba Wan3.0 Turns PDFs and Slide Decks Into 30-Second Videos

Alibaba's new video model starts from a document

Alibaba's Wan3.0 does something its video-generation rivals don't advertise: you can give it a business document and get a video back. The model belongs to the Wan family built by Tongyi Lab. It accepts the usual text, images, audio and video, and it also reads web pages, PDFs, PowerPoint files and spreadsheets, then produces a clip that follows the structure of the source.1314 A marketing or training team could start from a slide deck or a quarterly spreadsheet instead of writing a script and storyboard.1215

Outlets agree on the main specs. Wan3.0 makes clips of up to 30 seconds in a single pass at resolutions up to 1080p, which is double the 15-second limit of its predecessor, Wan2.7.1519 Alibaba says the model keeps characters, objects, backgrounds and spatial layouts consistent through a clip, and that it can produce synchronized facial expressions and speech in several languages. eWeek points out that nobody has independently verified those claims yet.15

The timing has been told in slightly different ways. A product write-up says Tongyi Lab opened a public beta on August 6, limited to invited accounts, under the model ID wan3.0-video.16 The wider rollout came on Monday, August 24, one day after Alibaba closed a HK$80 billion (about $10.2 billion) share placement in Hong Kong whose proceeds are meant to go to AI.18 Some coverage calls the August 24 step general availability.20 Other reports still describe the model as being in public beta.1417 On the evidence available, the safest reading is that access widened on August 24 but Alibaba is still positioning the model as an early product.

Why reading documents matters more than clip length

The input side is where Wan3.0 differs from its competitors. One comparison found that Sora 2 and Google's Veo 3.1 take text prompts and image references, and that Kling 3.0 and Seedance 2.5 also take video references. By that account, none of the four reads office files natively.16 The same analysis says Wan3.0 parses a document's headings, bullet points and table values itself, rather than treating the page as a flat image.16

Not every report agrees that Alibaba has this to itself. One developer-focused comparison lists Seedance 2.5 as having "limited" document input, not none.20 That suggests Wan3.0's lead may be a matter of degree. Even so, Alibaba is clearly the one making document-to-video the main selling point.

The workflow change is easy to see. With a prompt-only model, a team has to read the deck, summarize it, write a script, storyboard it and then generate the video, roughly five manual steps. Wan3.0 aims to cut that to uploading the file with a short prompt.16 People still have to judge the result: someone needs to check that the model kept the right points from a 40-slide deck.16 There are practical limits too. Each request takes one file or link, with a cap of 100MB and 50 pages.20

This is why most coverage treats Wan3.0 as a business productivity tool more than a creative toy. eWeek describes it as an opening into marketing, communications and training teams.15 Alibaba itself has pointed to training videos built from manuals, product demos made from slide decks, and video versions of quarterly reports.18 The company also mentions film production, short dramas and simulation footage for training self-driving and robotics systems.14 Reuters reported that the model had already been used during the beta for advertising, tourism promotion and music videos.15

Prices aimed below US rivals

Price is the other part of the pitch. List rates are $0.05 per second at 480p, $0.10 at 720p and $0.20 at 1080p, so a 30-second full-HD clip costs $6.14 A 30 percent launch discount brought that clip down to $4.20.14 The India Today report compares this with Google's Veo 3.1, whose standard tier costs $0.40 per second, double Wan3.0's top list rate.12 TechRound framed the $6 clip as a direct threat to human explainer-video work. It also argued that narrative design, brand voice and high-stakes investor communications will stay with people.17

There are trade-offs. Wan3.0 has closed weights, which breaks with Wan 2.1 and 2.2, both released openly.1316 Pre-launch talk of 4K output didn't hold up: the top resolution is 1080p.16 Speed may be the weakest point. In fal's published latency comparison, Wan3.0 averaged about 177 seconds per generation, against roughly 3.5 seconds for a rival called H3 Max, with similar preference ratings.20 That gap means little for a quarterly-report video. It matters a lot for any interactive or high-volume use.

How Wan fits into Alibaba's wider push

Wan3.0 is one of many Alibaba releases this year. Coverage linked it to the momentum of the Qwen language models, which Alibaba said had beaten Gemini and ChatGPT on coding benchmarks.12 Since the Wan launch, the Qwen side has kept shipping. Qwen3.8-Flash-Next arrived in late August as a preview of the architecture behind Qwen4.56 At the Apsara Conference on September 22, Alibaba said Qwen4 was in training. It did not give a release date.45 Most recently, on October 9, Qwen released Qwen-Image-2.1-Turbo, a 7B image model that cuts generation from 40 denoising steps to 8.28 Its weights are under a research license, so commercial self-hosting needs separate permission.28

The money explains much of the pace. The Decoder reported that Alibaba's quarterly profit fell 75 percent year over year as AI spending rose.14 At Apsara, Alibaba laid out a full-stack plan covering chips, cloud, models and agents.1 Wan3.0 is available through Alibaba Cloud Model Studio and Qwen Cloud, which suggests the video model is also meant to bring customers to Alibaba's cloud.14

Claude and Gemini releases keep coming

Wan3.0 arrived during an unusually busy run of releases from US labs. Anthropic released Claude Opus 5.5 on September 22, the first model in its 5.5 family.28 Claude Sonnet 5.5 followed on September 28 and Claude Haiku 5.5 on October 7. Haiku 5.5 is a fast model aimed at summaries, classification and customer support.2224 On Artificial Analysis's Intelligence Index, Opus 5.5 ranks first and Sonnet 5.5 second.28

Google announced Gemini 4 Argon on September 30, its first flagship since Gemini 3.1 Pro. It scored 52.6 on the same index and took first place on Arena's text and creative-writing boards.28 Access is still limited to partners.28 Google also shipped Gemini Nano Banana 2.1 on October 6.2224

The Claude and Gemini lines are also starting to overlap. A launch video for Google's new Gemini agent, published on October 8, shows a model picker listing Claude Opus 5 and Claude Sonnet 5.5 next to Gemini 4 Argon and Gemini 3.8 Flash.21 Google's release notes for its Antigravity developer tools name Claude Opus 5.5 and Sonnet 5.5.26 Claude access in the business agent is still listed as early access.21

Analysis: Alibaba is competing on a different axis

The US frontier race is mostly about reasoning benchmarks, agent features and per-token prices. Alibaba's video strategy looks deliberately different. It isn't chasing ever-longer clips, since Seedance 2.5 already matches the 30-second length.16 Instead it is betting on two things: which inputs the model accepts, and price. In my reading, that is a sound commercial bet. Companies already hold huge archives of decks and reports. A tool that turns that archive into usable video at a few dollars a clip addresses a real need, more than another cinematic demo would.

The caveats are real, though. Quality claims are unverified.15 Generation is slow.20 Closing the weights gives up some of the developer goodwill the earlier open Wan releases earned.16 Independent tests on real corporate documents, more than any leaderboard, will decide whether document-to-video becomes a standard feature that rivals copy or stays an interesting Alibaba-only option.