Wan AI: Alibaba's video model — what Wan 3.0 can do

What is Wan AI

Wan is Alibaba’s family of video models. It’s known for two tracks: open models you can download and run yourself, and cloud flagships.

Wan 2.1 and Wan 2.2 were released by Alibaba on GitHub and Hugging Face under the Apache 2.0 license. Wan 2.2 came out on July 28, 2025, with a Mixture-of-Experts architecture and models for text-to-video, image-to-video, speech-to-video and character animation. The smaller Wan 2.2 TI2V-5B generates 720p at 24 fps and runs on an RTX 4090-class GPU.

Wan 3.0 is the cloud flagship. Alibaba first opened a beta, then launched the API for everyone on August 24, 2026. A faster Wan 3.0 Prime came out alongside it. The previous Wan 2.7 made clips up to 15 seconds — Wan 3.0 doubled that.

What Wan 3.0 can do

30 seconds in one generation. Wan 3.0 makes up to 30 seconds at 30 fps and suggests a suitable duration based on the prompt. You can lay out several shots inside one clip, and the model will stitch them together with a shared soundtrack.

Sound and speech. Wan 3.0 generates character dialogue, scene sounds and background music without separate voiceover. Alibaba claims natural speech in several languages and a consistent voice for each character within a clip.

Up to 20 references. Besides text, the input can include images, videos, audio and — for the first time in Wan — documents and web pages: doc, xls, ppt, pdf, txt, key, pages, numbers, md, with one file up to 100 MB and 50 pages. Up to 5 audio clips totaling up to 15 seconds set the music or voice — that’s how you get dances on the beat and lip sync.

Editing without regenerating. In a finished clip, you can change the visuals, the plot or the dialogue without generating everything from scratch.

Vertical clips for social media and ads: 9:16 is set the same way as landscape.

Wan versions

	Wan 2.2	Wan 2.7	Wan 3.0
Released	July 28, 2025	before Wan 3.0	August 24, 2026
Duration	short clips	up to 15 s	up to 30 s
Resolution	480p, 720p	—	480p, 720p, 1080p
Sound	separate Speech-to-Video model	—	generated with the video
Open weights	yes, Apache 2.0	no	no
Arena rank (Sep 21, 2026)	42 (A14B)	—	6 of 48

What Wan is good for

• Short stories and Stories — 30 seconds and several shots in one generation.
• Clips from a document. Upload a presentation or a product description, and Wan 3.0 will pull the content from it.
• Dances and music videos — an audio reference sets the rhythm.
• Vertical ads with a product and sound.
• Your own infrastructure. The open Wan 2.1 and 2.2 can run on your own server with no API fees.

Wan vs. alternatives

	Wan 3.0	Seedance 2.5	Hailuo (MiniMax H3)	Gemini Omni
Developer	Alibaba	ByteDance	MiniMax	Google
Duration	up to 30 s	up to 30 s	4–15 s	3–10 s, extendable to 40
Resolution	up to 1080p	not disclosed by the developer	up to 2K	up to 4K
Sound	yes	yes	stereo	yes
References	up to 20, including documents	up to 50	up to 12	photos and videos
Arena rank	6	7 (2.5)	8	1–2

Wan 3.0 and Seedance 2.5 are the two models that make 30 seconds per generation. Wan stands out with document input and price, Seedance with the number of references. If you need 2K or 4K output, look at H3 and Gemini Omni.

How to write prompts for Wan

• Lay out the shots: “Shot 1: wide shot… Shot 2: medium… Shot 3: close-up…” — that way Wan won’t stretch one moment over 30 seconds.
• Put dialogue in quotes and say who’s speaking.
• Give sound its own line: “Sound: train noise, crunching snow, soft piano.”
• Match the duration to the content. Don’t ask for 30 seconds for a single action: Wan will suggest the right length itself.
• Keep on-screen text short. Alibaba admits that text in the frame isn’t always accurate yet.

How to use Wan in AI2Face

Wan is available in the AI2Face generator, no Alibaba Cloud account needed. You can run one prompt on Wan, Seedance, Hailuo and Gemini Omni and compare the results. The cost of each generation is shown before you run it, you can pay by card or crypto, and credits from packs don’t expire. New accounts get welcome credits after signing up.

Limitations

• Sound and text. Alibaba says outright that sound texture and on-screen text accuracy are still improving.
• Up to 1080p. Wan 3.0 has no 2K or 4K.
• 30 seconds isn’t cheap. A 1080p clip of that length costs $6 in the API — test the idea at 5–8 seconds in 480p first.
• Wan 3.0 is closed. Only Wan 2.1 and 2.2 can be downloaded.

Bottom line

Wan 3.0 makes some of the longest clips of any 2026 video model: 30 seconds with sound in one generation, 6th place on Arena and a rare ability to make a clip from a document. Use it for stories, Stories and ads where length and sound matter; for 2K, go with Hailuo H3, and for maximum quality, Gemini Omni. If you need your own infrastructure, the open Wan 2.1 and 2.2 are still the best free option. In AI2Face, Wan sits alongside all the other models.