Wan 3.0: 30-second videos from text, photos, audio and documents — what Alibaba's model can do

What is Wan 3.0

Wan 3.0 is the flagship video model in Alibaba’s Wan family. Alibaba opened a beta in early August 2026 and launched the API for everyone, with no application required, on August 24. An overview of the family, including the open Wan 2.1 and 2.2, is on the Wan page.

Alibaba calls Wan 3.0 an all-in-one model. A single model name covers text-to-video, frame-based generation, reference-based generation, editing and extension: Wan works out the task from the set of files and the wording of the prompt.

Wan 3.0 Prime, an accelerated version with the same features, came out alongside it.

On the Arena leaderboard as of September 21, 2026, Wan 3.0 ranks 6th of 48 video models. Only two versions of Gemini Omni, FLUX 3 Video from the FLUX 3 family, Grok Imagine Video 1.5 and Seedance 2.0 rank higher.

What Wan 3.0 can do

30 seconds in one generation. Length is 2 to 30 seconds at 30 frames per second. With “smart” duration, Wan picks a suitable length for the story itself. For multiple shots, Alibaba recommends 4–6 seconds per shot, laid out with timestamps.

Sound out of the box. Dialogue, background music and effects are generated together with the video. Audio is on by default and can be turned off.

Up to 20 references. Wan 3.0 accepts in one task:

Type	Limit
Images	up to 10, each up to 20 MB
Videos	up to 5, 15 s in total, each up to 100 MB
Audio	up to 5, 15 s in total, each up to 15 MB
Document or link	1 file (doc, docx, xls, xlsx, ppt, pptx, pdf, txt, md) or 1 public page

Video from a document. Upload a product deck or a report and briefly describe the tone — Wan 3.0 pulls the content from the file and builds the video. Alibaba’s examples: a deck becomes a brand film, a report becomes a narrated briefing. The file can be up to 100 MB and 50 pages.

Editing a finished clip. Upload a video and say what to change: “replace,” “remove,” “change to.” You can swap an object, the background or the dialogue without regenerating everything.

Extension. With “extend” or “continue,” Wan adds to a clip forward, backward or both. The source plus the extension can’t exceed 30 seconds.

First and last frame. One image sets the start; two set the start and the end.

Faces and emotion. Alibaba highlights diverse, realistic faces, micro-expressions and body language, plus crisp interfaces, charts and text — for app demos and video infographics.

Wan 3.0 vs Wan 3.0 Prime

	Wan 3.0	Wan 3.0 Prime
Purpose	main model	the same, faster
Length	2–30 s	2–30 s
Resolution	480p, 720p, 1080p	480p, 720p, 1080p
Audio	yes	yes
Price per second (480p / 720p / 1080p)	$0.05 / 0.10 / 0.20	$0.068 / 0.14 / 0.28
30 s at 1080p	$6.00	$8.40

Prime has the same features and parameters — Alibaba sells it as faster generation without a loss in quality. If you’re not in a hurry, regular Wan 3.0 is about 30% cheaper.

What Wan 3.0 is good for

• Short films, series and music videos — 30 seconds and several shots per generation.
• Ads from a brief — a product deck in, a video out.
• UI demos and infographics — Alibaba highlights crisp app screens and charts.
• Travel and destination films without a location shoot.
• Music-driven clips — reference audio sets the rhythm and voice.
• Editing and extension of finished clips without full regeneration.

Wan 3.0 vs alternatives

	Wan 3.0	Seedance 2.5	MiniMax H3	HappyHorse 1.1
Developer	Alibaba	ByteDance Seed	MiniMax	Alibaba
Length	2–30 s	4–30 s	4–15 s	3–15 s
Resolution	up to 1080p	up to 1080p	up to 2K	up to 1080p
References	10 photos, 5 videos, 5 audio	up to 50 files	9 photos, 3 videos, 3 audio	up to 9 photos
Document as input	yes	no	no	no
Editing and extension	yes	yes	editing	editing (1.0 model)
Arena rank (Sep 21, 2026)	6	7	8	10 (1.0)
Price per second, 1080p	$0.20	—	$0.13 (2K)	—

Wan 3.0 and Seedance 2.5 from the Seedance family are the two models that make 30 seconds per generation: Wan is cheaper and reads documents, while Seedance takes more references. MiniMax H3 wins on resolution, and HappyHorse is Alibaba’s second video model for short clips with dialogue.

How to write prompts for Wan 3.0

1. Refer to files by number: “Image 1,” “Video 1,” “Audio 1.” Each type has its own numbering, in upload order.
2. Lay out shots with timestamps, 4–6 seconds per shot.
3. Editing and extension need keywords: “edit video,” “replace,” “remove,” “change to” or “extend,” “continue.” Without them, Wan won’t understand the task.
4. With a document, keep the prompt short: the tone and goal — the model takes the content from the file.
5. Draft at 480p. Alibaba itself recommends testing the idea at low resolution and only then rendering 1080p.

An example based on the storyboard above:

0–6 s: wide shot, a grandfather in a flat cap and a girl with two braids climb a grassy
seaside hill with a red kite. 6–12 s: close-up, the grandfather ties the string and says:
"Hold on tight." 12–18 s: medium shot, the girl runs, the kite lifts off. 18–24 s:
low angle, the red kite high in the clouds. 24–30 s: wide shot, both sit on the grass
watching the sea at sunset. Sound: wind, surf, the girl laughing, a warm guitar melody.

How to use Wan 3.0 in AI2Face

Wan 3.0 is in the AI2Face generator next to MiniMax H3, Kling 3.0 and Veo 3.1.

1. Sign up — new accounts get starter credits.
2. Open the generator, switch to video and pick Wan 3.0.
3. Upload frames, references or a document, write a prompt, set the length and resolution.
4. The credit cost is shown before you start. Credits don’t expire.

A product-spec slide or storyboard frames are easy to make in the same generator — with GPT Image 2.5 or Nano Banana Pro.

Limitations

• Up to 1080p. Wan 3.0 has no 2K or 4K — for those, look at MiniMax H3 and Kling 3.0.
• Audio and text are still improving. Alibaba says outright that audio texture and on-screen text accuracy are being worked on.
• Modes can’t be mixed. References, a document or a link can’t be combined with first and last frames in one task.
• One document or one link — not both; the page must open without a login.
• Long clips are expensive. 30 seconds at 1080p costs $6 in the API, or $8.40 with Prime.
• No weights. Wan 3.0 is cloud-only; only Wan 2.1 and 2.2 can be downloaded.

Bottom line

Wan 3.0 is the most flexible long-form video model of 2026: 30 seconds with sound, 20 references, a video from a deck or a link, and editing and extension in one model. It’s 6th on Arena at $0.20 per second of 1080p — cheaper than most of its neighbors in the ranking. Use Wan 3.0 for stories, ads from a brief and music-driven clips; for 2K, go with MiniMax H3, and for dozens of references, Seedance 2.5. Prime is only worth it when speed matters more than price.