Wan 3.0: the first video model that can read a deck

What Wan 3.0 is

Wan 3.0 is a video generation model from Alibaba's Tongyi Lab, moved to general availability on 24 August 2026. On its own product feed, Alibaba Cloud announced it as the video generation model Wan 3.0 going live; the same day, Beijing News summarised it as 30 seconds per clip, and a PPT can become video too.

It is the third mainline generation of the Wan family. The first two — 2.1 and 2.2 — are among the most widely used video models in the open-weights community. Wan 3.0 does something that family had never done before: it treats a document as an input modality. Not "look at your deck and write a prompt about it", but hand the model the doc, xls, ppt, pdf or md file and let it read the structure and the text itself.

Put against its rivals, the position is clear. Through the second half of 2026, a native 30-second single generation has become the shared bar at the top of this category — ByteDance's Seedance 2.5, Kuaishou's Kling 4.0 and Alibaba's Wan 3.0 are all competing for it. What actually distinguishes Wan 3.0 is two other things: it reads documents, and its commercial positioning is written on its face — film and short-drama production teams, creator platforms, ad agencies, game and design studios, tourism promotion, music and entertainment. These are organisations that need to output at volume, not individuals.

What changed from Wan 2.x

Two rows deserve to be read twice. The first row is a change of identity: Wan 2.2 released its code and weights under Apache-2.0 on 28 July 2025, and most of the cheap local video workflows in the community are built on it. Wan 3.0 makes no statement about releasing anything, which means that line of workflows stops here. The second is sound: with Wan 2.x you generally generated the picture and dubbed afterwards. Native audio-video generation removes exactly that step, and with it the lip-sync rework.

What document input really does

This is Wan 3.0's only exclusive, so it is worth being precise about the edges. The published rules are clear: one file or link per request, up to 100MB and up to 50 pages, in doc, xls, ppt, pdf, txt, key, pages, numbers and md formats. Alibaba's own four examples:

• A product deck becomes a brand or launch film. Upload a brand-introduction document, add one sentence, get a finished brand film.
• A training deck becomes video courseware. Turning lecture material into something with pictures, rather than reading text off slides.
• Spreadsheet data becomes animated charts. Numbers turned into moving visualisation — the category most dependent on on-screen text being accurate.
• A business report becomes a narrated briefing. Conclusions turned into a voiced film, for internal review or external communication.

On mechanism, Alibaba's explanation is prompt engineering and instruction following: the model treats the prompt as the control centre, and that is what interprets the input, orchestrates the content and shapes the final expression. Translated into practical terms: the document is the material, the prompt is still the director. Dropping in a deck with no prompt gives the model no direction.

Where you can actually use it

Like Seedance, Wan 3.0 is not a product you sign up for once — it is a model that reaches you through several different routes, and the route decides your price, your limits and your rights.

• The Wanxiang site — Tongyi's own consumer entry point. The place to start if you want to see what it does without touching an API.
• Alibaba Cloud Bailian — the mainland developer route. The API is named wan3.0-video, and Alibaba specifies it can be called without an application, billed per second, with a Prime accelerated-inference service alongside. This is the most direct way to put it inside your own product.
• The Qianwen AI platform and app — one of the first routes Alibaba published, aimed at people already using Qianwen.
• Third-party platforms — Alibaba states Wan 3.0 has gone live on Flova, Libtv, Jurilu and Haiyi. Convenient, but you are paying a markup and the terms are not necessarily Alibaba's terms.

Three practical consequences. First, confirm on the interface whether video editing is exposed: Alibaba says the capability carries forward from Wan 2.7, but individual platforms do not all surface the same features. Second, a platform's credits are not a unit of video — any table converting credits straight into clip counts without stating resolution and duration is a guess. Third, failed generations are not billed only where Alibaba says so; whether a reseller follows that is a separate question to ask.

What each tier costs

The pricing model is clean: per second of generated video, with resolution setting the rate. These are the figures Alibaba published, read in October 2026.

Run the common case at 720P: a 30-second finished piece costs 18 CNY for the generation itself. Apply the two-to-three re-rolls from earlier and the real cost lands between 36 and 54 CNY per usable clip. At 1080P, double it. That is what per-second billing means in practice — duration is the first-order cost variable, and re-roll count is the second, and the second one usually costs more.

The three things it does not say

This section is where the page differs from most of what gets republished. We went looking for three things. Here is what we found.

None of this means the model is weak — it is unambiguously the most direct route from office material to video, and that is not in dispute. But it does mean that any claim you read about Wan 3.0's parameter count, training data or superiority over a named rival should be treated as marketing until it has a primary source.

Getting a usable clip out of it

Document input changes one thing: your starting point used to be a prompt, and now it can be a file. It also introduces a new way to fail.

1. Edit the document before you upload it.

   The model reads the structure and the text, so a 50-page deck that covers everything hands it 50 conflicting directions. Cut down to the three pages you actually want to communicate and put the conclusion in one sentence on the first one. Alibaba says the prompt is the control centre — so give the control centre a target.

2. Write the prompt as a timetable, not as adjectives.

   Thirty seconds means the model can read integer-second structure. "0–6s: product close-up, slow push in. 6–18s: in use. 18–30s: logo and end card." That is what makes it controllable rather than lucky, and it is the most substantive upgrade over the previous generation.

3. Rehearse at low resolution, render only the final at high.

   The rate steps up with resolution. Lock the 30-second rhythm and framing at 480P, then re-render only the take you are keeping at 1080P. This is your largest cost lever — far larger than shopping around for a cheaper platform.

4. Edit on the timeline before you re-roll.

   Alibaba keeps video editing in the release, with visuals, plot and dialogue all editable. Its whole point is not throwing away twenty-nine good seconds because of one bad one. Think edit first, regenerate second.

5. Pull out every shot with text in it and check it by hand.

   Alibaba admits on-screen text rendering is still improving, and in document-to-video work the on-screen text often is the information. Titles, figures, logos, chart labels — each is worth its own look. This one is specific to Wan 3.0.

6. Label it, and keep a record of where your references came from.

   Publish with the platform's AI-generated marker, and log the source of every reference asset and document. Wan 3.0 is positioned as enterprise-grade volume production; clients will ask, and this record is what protects you when they do.

Frequently asked questions

What is Wan 3.0?

Alibaba Tongyi Lab's video generation model, moved to general availability on 24 August 2026. It is the first generation of Wan to reach a 30-second single generation and the first to accept document input — doc, xls, ppt, pdf and md files handed over as source material, which Alibaba describes as turning anything into video. Routes include the Wanxiang site, Alibaba Cloud Bailian, the Qianwen AI platform and the Qianwen app; the API model name is wan3.0-video.

How much does Wan 3.0 cost?

Per second of generated video, with resolution setting the rate. On the mainland platform, 480P / 720P / 1080P are 0.3 / 0.6 / 1.2 CNY per second; internationally the same tiers are $0.05 / $0.10 / $0.20 per second. The API needs no application and failed generations are not billed. The launch discount ended on 23 September 2026 — the 0.21 CNY figure still circulating is the old price.

Is Wan 3.0 open source? Can I run it locally?

No. Wan 2.2 released code and weights under Apache-2.0 on 28 July 2025 (Hugging Face and ModelScope under Wan-AI, plus GitHub), but Wan 3.0's announcement covers APIs only. We checked the Wan-AI organisation on Hugging Face — 27 repositories, topping out at the Wan 2.2 family, no Wan 3.0. So there is no legitimate local route. If offline generation is a hard requirement, plan around open-weight models like Wan 2.2 instead — see the tools comparison.

Does document input actually work?

The capability is published and so are the limits: one file or link per request, up to 100MB and 50 pages. Alibaba's examples are product decks to brand films, training decks to courseware, spreadsheets to animated charts and reports to narrated briefings. But the same FAQ admits on-screen text rendering is still improving, and these tasks depend on exactly that. Use it as an accelerator for shot planning and visuals, not as a one-click deck-to-film tool.

How long can a clip be?

Up to 30 seconds in a single generation, up to 1080P, with native audio-video generation. Smart duration recommendation reads your prompt before suggesting a length, and video extension continues a story past the first generation. Video editing carried over from Wan 2.7 also lets you modify visuals, plot and dialogue without regenerating the clip.

What is the difference between Wan 2.2 and Wan 3.0?

Duration goes from short clips to 30 seconds, with smart duration and extension added. Document and web-page input arrive. Realism is called out as a headline upgrade — Alibaba's phrasing is "a thousand faces for a thousand people", meaning no more cookie-cutter AI faces. And one more that gets overlooked: Wan 2.2 is open source, Wan 3.0 is not.

Is Wan 3.0 safe for commercial projects?

Alibaba names its target customers as film and short-drama production, creator platforms, advertising and marketing, game and design studios, tourism promotion and music entertainment — organisations that output at volume — and says it was already running in their production pipelines during the preview period. But usable commercially is not the same as safe for anything: reference photos of real people, protected brand assets and other people's work remain your risk. Use original subjects and keep a source record.

How does it compare with Seedance 2.5 and Kling 4.0?

All three are chasing the same 30-second single-generation bar with different emphases. Wan 3.0's exclusive is document input, making it the most direct route from office material to video. Seedance 2.5 is audio-video native, strong on synchronised speech and on combining up to 50 reference assets. Kling 4.0 is, as of October 2026, still announced but not fully available. There is no like-for-like public test between them — judge on your own footage.

Also on this site

• Seedance 2.5 — ByteDance's 30-second model on the audio-video native route, and the one with the largest reference budget of the three.
• Kling 4.0 — the announced-but-not-yet-shipped generation from Kuaishou: what the company actually said, what is only reported, and what the credits really cost.
• MiniMax H3 — first place on Arena's image-to-video blind test, and one of the few frontier video models whose weights you can download. A direct contrast with Wan 3.0's closed release.
• Higgsfield Genjutsu — recast a video you already shot: change the character, the location or an object without a reshoot, and without writing a prompt.
• Free AI video tools compared — which category of tool is genuinely free, what each one reuses from your source material, and what to verify before you pay.