FrameSurfer publishes this guide and is one of the tools covered.
Wan is Alibaba's family of AI video models. Wan 2.5 makes 5 or 10 second clips with sound, at up to 1080p. Wan 3.0, launched on August 24, 2026, makes up to 30 seconds in one generation, also at up to 1080p (Alibaba Cloud). Only the older Wan 2.1 and 2.2 are open source. Wan 2.5 and newer have no public weights, so you use them through an API or an app, as of September 2026.
This guide compares the versions, shows what each one costs, and explains when to run Wan yourself and when to use an app.
The Wan family at a glance
Here is how the versions line up, based on Alibaba Cloud's video generation model list and Wan's official Hugging Face and GitHub pages.
| Version | Open weights | Clip length | Resolution | Sound |
|---|---|---|---|---|
| Wan 2.1 (February 2025) | Yes, Apache 2.0 | 5 seconds on the API | 480p, 720p | No |
| Wan 2.2 (July 2025) | Yes, Apache 2.0 | 5 seconds on the API | 480p and 720p open models, up to 1080p on the API | No |
| Wan 2.5 (preview) | No | 5 or 10 seconds | 480p, 720p, 1080p | Yes |
| Wan 2.6 | No | 2 to 15 seconds | 720p, 1080p | Yes |
| Wan 2.7 | No | 2 to 15 seconds | 720p, 1080p | Yes |
| Wan 3.0 (August 2026) | No | 2 to 30 seconds | 480p, 720p, 1080p | Yes, with an on/off switch |
"Open weights" means you can download the trained model files and run them on your own computer. Apache 2.0 is a permissive license that allows commercial use.
Is Wan open source?
Some of it. The short answer by version:
- Wan 2.1 and Wan 2.2: open. Their weights are on the Wan-AI Hugging Face page and their code is on GitHub, under Apache 2.0.
- Wan 2.5, 2.6 and 2.7: not open. On September 24, 2026, the Wan-AI Hugging Face page had no weights for them.
- Wan 3.0: not open. The official Wan 3.0 GitHub repository holds only a README and a license file, with no model files or code. Its README still uses the phrase "open-source video synthesis", which confuses people.
Alibaba's open releases in 2026 have been specialist models instead. In July 2026 the Wan-AI page added Wan-Dancer-14B and Wan2.2-Animate-2, not new general text-to-video versions.
What it takes to run Wan on your own GPU
Wan's own READMEs list the hardware. VRAM is the memory on your graphics card, and it decides which model you can load.
| Open model | What it does | Hardware Wan lists |
|---|---|---|
| Wan 2.1 T2V-1.3B | Text to video, best at 480p | 8.19 GB of VRAM. About 4 minutes for a 5-second 480p clip on an RTX 4090, per Wan 2.1's README. |
| Wan 2.2 TI2V-5B | Text or image to video, 720p at 24 fps | At least 24 GB of VRAM, such as an RTX 4090. Under 9 minutes for a 5-second 720p clip on one consumer GPU, per Wan 2.2's README. |
| Wan 2.2 T2V-A14B and I2V-A14B | Text or image to video, 480p and 720p | At least 80 GB of VRAM on a single GPU |
None of the open Wan models generate sound. Wan2.2-S2V-14B is a speech-to-video model, which takes audio as an input rather than creating it.
Wan 2.5 vs Wan 3: what changed
Wan 2.5 added sound. Wan 3.0 added length, references and editing. Here are the two side by side, from Alibaba Cloud's model docs and the Wan 3.0 launch post.
| Wan 2.5 | Wan 3.0 | |
|---|---|---|
| Clip length | 5 or 10 seconds | 2 to 30 seconds, with a suggested length based on your prompt |
| Resolution | 480p, 720p, 1080p | 480p, 720p, 1080p |
| Frame rate | 30 fps | 30 fps |
| Sound | Generates matching sound, or syncs to an audio file you supply | Native audio with an on/off switch, plus reference audio |
| Starting points | Text, or a first image | Text, first and last frames, images, video, audio, documents and web pages |
| References | Not listed | Up to 10 images, 5 videos and 5 audio clips |
| Editing | No | Change visuals, plot and dialogue without starting over |
| API price at 1080p | $0.15 per second | $0.20 per second |
Two things stand out.
First, the document input is new. Wan 3.0 reads one file or link per request, such as a PDF, a slide deck or a spreadsheet, up to 100 MB and 50 pages. Alibaba's examples include turning a product deck into a launch ad and a report into a narrated briefing.
Second, Wan 3.0 costs the same per second as Wan 2.5 at 480p and 720p. It only costs more at 1080p. So for most short-form work, the newer model does not cost more per second.
Alibaba is also open about the gaps. Its launch post says audio texture and on-screen text accuracy in Wan 3.0 are still improving.
The versions in between built toward this. Wan 2.6 moved to 2 to 15 second clips with multi-shot scenes. Wan 2.7 offered first and last frame control and video continuation, and introduced the video editing that Wan 3.0 carries forward.
Wan API pricing on Alibaba Cloud Model Studio (as of September 2026)
Model Studio is Alibaba Cloud's platform for calling its AI models by API. It bills per second of finished video, and failed requests cost nothing. These are the international (Singapore) text-to-video prices on the Model Studio pricing page, as of September 2026.
| Model ID | 480p | 720p | 1080p |
|---|---|---|---|
| wan3.0-video | $0.05 per second | $0.10 per second | $0.20 per second |
| wan3.0-video-prime (faster) | $0.068 per second | $0.14 per second | $0.28 per second |
| wan2.7-t2v | Not offered | $0.10 per second | $0.15 per second |
| wan2.6-t2v | Not offered | $0.10 per second | $0.15 per second |
| wan2.5-t2v-preview | $0.05 per second | $0.10 per second | $0.15 per second |
| wan2.2-t2v-plus | $0.02 per second | Not offered | $0.10 per second |
What that means per clip:
- A 10-second Wan 2.5 clip at 720p costs $1.00.
- A 30-second Wan 3.0 clip costs $1.50 at 480p, $3.00 at 720p and $6.00 at 1080p.
A few more rules from the same page:
- Wan 3.0 bills input video too. When you send in a video to edit or reference, its seconds are added to the bill.
- Discount note. The wan3.0-video rows showed a "limited-time 30% off" note next to the list price.
- Free quota. Model Studio lists a free quota of 30 seconds for Wan 3.0 and 50 seconds for each Wan 2.x model, valid for 90 days. It applies only in the Singapore region.
- Region matters. Prices differ by region. The China (Beijing) rates, for example, are lower.
Wan app plans (create.wan.video)
If you would rather not touch an API, Alibaba runs its own Wan app on the web and on mobile. These are the plans the Wan pricing page showed on September 24, 2026.
| Plan | Billed monthly | Billed yearly | Credits per month | What the page lists |
|---|---|---|---|---|
| Free | $0 | $0 | Daily check-in credits | 1 video at a time |
| Pro | $6.50 | $5 a month | 300 | Up to 60 videos, 1080p, 10 to 30 second videos, watermark-free downloads |
| Premium | $26 | $20 a month | 1,200 | Up to 240 videos, 1080p, 10 to 30 second videos, watermark-free downloads |
The page showed $10 and $40 as crossed-out list prices, so the monthly prices were discounts. It also says app plans cover create.wan.video only. The API needs a separate purchase. The first month or year is not refundable once active.
Is Wan AI free?
Yes, in two ways. The Wan app has a free plan with daily check-in credits, though 1080p, longer videos and watermark-free downloads are listed on the paid plans. And the Wan 2.1 and 2.2 weights are free to download, if you have the hardware.
Running Wan yourself vs using an app
| Route | What you pay | Best for | Trade-offs |
|---|---|---|---|
| Your own GPU | Nothing for the weights, but you pay for the hardware | Tinkerers, private data, custom training | Wan 2.1 and 2.2 only. No sound. Slow on home GPUs. |
| Model Studio API | Per second of video | Developers building a product | You build editing, captions and voiceover yourself. |
| Wan app | $0 to $26 a month | Trying Wan 3.0 on its own | 1080p and longer videos need a paid plan. App credits do not cover the API. |
| FrameSurfer | One plan for more than a dozen models | Finished short videos | Wan 3 starts from text or a start image. No document or multi-reference inputs. |
On FrameSurfer, Alibaba Wan 3 makes clips of 2 to 30 seconds at 480p, 720p or 1080p, with native audio. It starts from text or a start image. You can also set the aspect ratio, a seed, prompt expansion and a negative prompt. A negative prompt lists what you do not want, such as "blurry" or "text on screen". Wan 3 makes the longest single clip of the more than a dozen models FrameSurfer runs. That helps with one-take shots, like a slow walk-through or a product reveal that should not cut.
FrameSurfer then builds the rest of the video. Give it an idea, a photo, a script or a link, and it writes the script, generates each shot, and adds voiceover, captions and music. You refine it in a visual editor and export an MP4 at up to 1080p for TikTok, Reels, Shorts or ads.
To try a 30-second Wan 3 shot, open FrameSurfer's text to video tool, set the length, and pick the model.
Other open source AI video generators
If "open source" is your main filter, Wan 2.1 and 2.2 are only part of the story.
- MiniMax H3, the model behind Hailuo AI, has open weights under its own community license, with extra terms for companies above $20 million in yearly revenue. It is much larger than the open Wan models. See the Hailuo AI video guide.
- Open weights are not the same as free output. You still pay for GPUs or cloud time, and you build the editing, sound and captions yourself.
For closed models with native audio, compare Seedance 2.0 pricing and Veo 3 pricing. For every model FrameSurfer runs, side by side, see AI video models compared. For whole tools rather than models, see the best AI video generators.
Have a photo you want to bring to life? FrameSurfer's image to video tool makes it the first frame and animates it with Wan 3 or any other model on your plan.
Frequently Asked Questions
Is Wan 2.5 open source?
No. As of September 24, 2026, Alibaba has not published Wan 2.5 weights on its official Wan Hugging Face or GitHub pages. Wan 2.1 and Wan 2.2 are the open versions, under Apache 2.0.
Is Wan AI free?
The Wan app at create.wan.video has a free plan with daily check-in credits. The open Wan 2.1 and 2.2 models are free to download, but you need your own GPU to run them.
What is the difference between Wan 2.5 and Wan 3?
Wan 2.5 makes 5 or 10 second clips with sound. Wan 3.0 makes 2 to 30 second clips and adds references, document input and video editing. Both reach 1080p.
How long can a Wan 3 video be?
Up to 30 seconds in a single generation, at 480p, 720p or 1080p. Alibaba says Wan 3.0 also supports video extension, which continues an existing clip.
Can I run Wan on my own computer?
Yes, for Wan 2.1 and 2.2. The small Wan 2.1 model needs about 8 GB of VRAM, and Wan 2.2's 5B model needs at least 24 GB. Wan 2.5 and newer cannot be run locally because their weights are not public.
Who makes Wan AI?
Alibaba makes Wan. The Wan website says it is powered by Alibaba Cloud, and the API runs on Alibaba Cloud Model Studio.
