Alibaba's open-source AI video model is called Wan 2.2. It was released on 28 July 2025 (Apache 2.0 license) and will convert a text prompt or still image into a short clip (720p). It is built with a “Mixture-of-Experts” design and can be deployed on dedicated or custom hardware.
It used to be the latest Wan model to download more than a year ago. Since then, Alibaba has released Wan 2.5, 2.6, 2.7, and 3.0, all of which are hosted services and don't have any public weights. In this guide, you will read about what Wan 2.2 is, which variant to use, what hardware is required, how to use it, and some comparisons to Wan 3.0.
Wan 2.2 at a glance
Released: 28 July 2025, by Alibaba’s Tongyi/Wan team.
License: Apache 2.0. Code and weights are public, and commercial use is allowed.
Design: Mixture-of-Experts diffusion model. 27B parameters in total, 14B active per step.
Inputs: text-to-video, image-to-video, a hybrid text-and-image model, plus speech-to-video and character animation variants.
Output: short clips up to 720p. The small hybrid model does 720p at 24 fps.
Tools: ComfyUI workflows, the official GitHub repo and Hugging Face weights.
Newest Wan: Wan 3.0, which is hosted and, as far as we can confirm, has no public weights.
Wan 2.2 is no longer the newest Tool
Alibaba has kept the Wan name moving. Wan 2.5 arrived as a preview in September 2025, Wan 2.6 in December 2025 and Wan 2.7 in April 2026. Wan 3.0 went into public beta on 6 August 2026, with the formal release on 24 August. It runs as an API on Alibaba Cloud’s Model Studio, makes clips of up to about 30 seconds, and bills per second of video.
None of those four published their core weights. Wan 2.1 and Wan 2.2 are the open ones. Alibaba has released a few side models under Apache 2.0, like the music-to-dance model Wan-Dancer-14B and the character-animation model Wan-Animate-2, but those sit outside the main video line.
Some sites claim Wan 3.0 or 2.7 is open source. Several independent trackers looked for the weights on Alibaba’s official Hugging Face, GitHub and ModelScope pages and found nothing. So before you download anything labeled Wan 3.0, check whether it comes from the official Wan-AI organization.
What is Wan 2.2?
Wan 2.2 is the second open generation of Alibaba’s Wan video family, built by the Tongyi team. Most strong video models sit behind an API. This one releases its code, weights and architecture, allowing researchers, startups and single creators to run it locally and develop on it.
There were three major differences from Wan 2.1: the introduction of a Mixture-of-Experts architecture, increased control over the appearance of a shot, and a significantly larger training set. Alibaba's materials include the elements of lighting, composition, color tone, and camera movement.
How does the Mixture-of-Experts design work?
A diffusion model starts from noise and cleans it up over many steps. Wan 2.2 splits that work between two experts. A high-noise expert takes the early steps, where the overall layout and motion are decided. A low-noise expert takes over later and refines the detail. Only one runs at a time, so the model holds 27B parameters but uses about 14B per step.
The payoff is more capacity without a matching jump in computer. The cost per step stays close to a 14B model. MoE is common in language models, but it was an unusual choice for open video generation when Wan 2.2 launched.
Which Wan 2.2 variant should you use?
Variant | Input | Size | Best for |
T2V-A14B | Text to video | MoE, 27B total / 14B active | Best quality from a text prompt |
I2V-A14B | Image to video | MoE, 27B total / 14B active | Animating a still image |
TI2V-5B | Text or image to video | 5B | Local use on one consumer GPU, 720p at 24 fps |
S2V-14B | Speech (audio) to video | 14B | Talking or singing characters driven by audio |
Animate-14B | Character animation and replacement | 14B | Moving a character using a reference video |
New to this? Start with TI2V-5B. It uses a high-compression VAE, so it fits on a single consumer card, and it takes both text and image input. Move up to the 14B models once you need finer detail and have the memory for it.
What hardware do you need to run Wan 2.2?
The 5B hybrid model is built for one consumer GPU with around 24 GB of VRAM, like an RTX 4090. The 14B models need far more. The official repo documents ways to cut that, such as offloading parts of the model to the CPU, and the community has published lighter builds. Check the README for your variant, since the real requirement depends on resolution and settings.
You’ll also want fast storage for the weights and enough system RAM for pre- and post-processing. And don’t expect instant results. On consumer hardware, a clip takes minutes, not seconds.
How to install and run Wan 2.2 locally
Copy the repo from github.com/Wan-Video/Wan2.2
Install Python and the required dependencies (from the repo), as well as a compatible PyTorch version
Get the weights of your selected variant from the official Hugging Face organization of the Wan-AI organization
Execute the inference script using a prompt, an image, or both, in the same way as in the examples given in the “README” section
Follow precisely as stated in the “README”. The flags and options may change with releases, so the content here may be outdated a month later.
Can you use Wan 2.2 in ComfyUI?
Yes, and for most creators it’s the easier way in. Wan 2.2 launched with day-zero ComfyUI support and ready-made workflows for text-to-video, image-to-video and the hybrid model. You load the model, type your prompt, and connect the decoder and video output nodes. No scripting.
Is Wan 2.2 free for commercial use?
The code and weights are under Apache 2.0, which permits commercial use. Running it yourself costs hardware and electricity, not a license fee. Some online tools offer Wan 2.2 through their own interfaces, and those may charge. Read the license text in the repo, and remember that you’re responsible for what you generate, including copyright, likeness and disclosure rules.
Wan 2.2 vs Wan 3.0
Particulars | Wan 2.2 | Wan 3.0 |
Released | 28 July 2025 | Public beta 6 August 2026, formal release 24 August 2026 |
Weights | Public, Apache 2.0 | No public weights confirmed |
How you use it | Run it yourself or in tools like ComfyUI | Alibaba Cloud Model Studio and partner platforms, hosted API |
Clip length | Short clips, typically a few seconds | Up to about 30 seconds |
Cost model | Your own hardware | Billed per second of video |
Best for | Local control, privacy, customization, fine-tuning | Longer clips and the latest hosted quality |
The trade-off is simple. If you want your prompts and outputs to stay on your own machine, want to fine-tune, or rather skip per-second fees, pick Wan 2.2. If you need longer, more polished clips and a hosted API is fine, Wan 3.0 would be the best choice.
How does Wan 2.2 compare with closed video models?
Alibaba reports strong results for Wan 2.2 on its own Wan-Bench evaluation. That’s the vendor’s benchmark, so treat it as a starting point, not proof. Where the open model really wins is transparency, control and cost. You can inspect it, fine-tune it, run it offline and build it into a product without per-clip fees.
Hosted closed models usually win on convenience, resolution options and clip length, and they handle the infrastructure for you. Several newer ones, including Alibaba’s own later Wan versions, are built for longer and more polished output. If one type of shot matters to you, run the same prompt on both and compare before you commit.
What can you build with Wan 2.2?
Think short promo clips and social content, animated versions of product photos or illustrations, storyboard and concept previews, motion tests for games, and educational animations. Because it’s open, it also fits workflows where cloud fees or data privacy rule out hosted tools. Many people fine-tune it with LoRA-style methods to teach it a specific look.
What are the limitations?
Short clips: One generation is a few seconds, so anything longer means stitching clips together.
Resolution: Output tops out around 720p, which is below what some hosted models offer.
Hardware: The 14B models need far more memory than most home setups have.
Consistency: Flicker, drift, and odd motion still happen, like with every video model.
Misuse risk: Open video models can be abused for deepfakes and misinformation. Get consent, label your output, and set sensible policies.
Not the frontier anymore: Alibaba’s newer hosted models are ahead.
Final verdict
Wan 2.2 isn’t the shiny new thing anymore. But it does something none of Alibaba’s newer models do: you can download it, run it, inspect it, and fine-tune it. If you care about privacy, cost control, or customization, that still counts for a lot. For the longest, most polished clips, look at hosted models, Wan 3.0 included. And if a site tells you there’s a newer open Wan, check the official Wan-AI pages first.
Want help building AI video or other generative AI into your product or workflow? Our AI Development team builds it, and our AI Consulting team can help you choose the right models first. Book a free consultation.
For more on fast-moving AI models, read our Gemini 3.7 Flash guide and The Future of Generative AI: 8 Key Trends.
Frequently asked questions
When was Wan 2.2 released?
28 July 2025, by Alibaba’s Tongyi/Wan team.
Is Wan 2.2 open source?
Yes. The code and weights are public under the Apache 2.0 license.
Is Wan 2.2 free?
Downloading and running it on your own hardware is free. Online tools that host it may charge for usage.
Is Wan 2.2 the latest Wan model?
No. Alibaba has since released Wan 2.5, 2.6, 2.7 and 3.0. But Wan 2.2 is still the newest main Wan model with public weights.
What is the difference between Wan 2.2 and Wan 3.0?
Wan 3.0 is newer and hosted. It supports clips of up to about 30 seconds and bills per second. Wan 2.2 is open, so you can run and customize it yourself.
What are the Wan 2.2 hardware requirements?
The 5B hybrid model targets a single consumer GPU with around 24 GB of VRAM, like an RTX 4090. The 14B models need much more. Check the official README for current guidance.
How long are Wan 2.2 videos?
A few seconds per generation, at up to 720p. For longer videos, you stitch clips together.
Does Wan 2.2 support audio or speech?
The main text and image models make video only. A separate speech-to-video variant, S2V-14B, generates video driven by audio input.
Can I use Wan 2.2 in ComfyUI?
Yes. It has official ComfyUI support, with ready-made workflows for each main variant.
Can I use Wan 2.2 commercially?
The Apache 2.0 license allows it. You’re still responsible for how you use the generated content.