Updated: September 3, 2026 · Docs · Clothoff AI Editorial Team
What is Wan (Alibaba)?
A large video family that ships open weights next to a closed API line — which is why it appears wherever hosted models say no.
Wan is a family of video generation models developed by Alibaba and released from 2025 onward in two lines: open-weight checkpoints such as Wan2.2 under the Apache 2.0 license, and closed, API-only releases such as Wan2.7 and Wan3.0 served through Alibaba Cloud Model Studio. It covers text-to-video, image-to-video and reference-to-video generation.
| Version | Date | Developer | License / Access | Filters |
|---|---|---|---|---|
| Wan2.1 | 2025 (day not stated on the repository) | Alibaba (Wan team) | Open weights on GitHub and Hugging Face | No safety filter stated |
| Wan2.2 (T2V/I2V 14B, TI2V 5B) | Jul 28, 2025 | Alibaba (Wan team) | Apache 2.0 open weights | No built-in filter; hosts add their own moderation |
| Wan2.7 (t2v, i2v, r2v, videoedit) | Apr 3, 2026 | Alibaba Cloud | API-only via Model Studio | Provider-side; NSFW rules not stated |
| Wan3.0 (wan3.0-video, -prime) | Aug 6, 2026; general availability Aug 24, 2026 | Alibaba Cloud | API-only, billed per generated second | Not stated on the GA announcement |
Dates are the earliest listing confirmed on Alibaba or repository pages; Model Studio dates follow the Singapore region list.
Which Wan versions are open weights and which are API-only?
Wan2.1 and Wan2.2 are downloadable checkpoints published on GitHub and Hugging Face. Wan2.2 ships as three variants: text-to-video and image-to-video models with 14B active parameters each, plus a hybrid 5B model for both tasks, all under Apache 2.0.
Everything after Wan2.2 is closed. Wan2.7 appeared in the Model Studio release notes on Apr 3, 2026 with text-, image-, reference-to-video and video-edit endpoints; Wan3.0 followed on Aug 6, 2026 and reached general availability on Aug 24, 2026.
“Wan3.0 is available through Alibaba Cloud Model Studio and supported international channels, with API pricing billed per generated second.”
— Alibaba Cloud Community blog, accessed Sep 3, 2026
Does Wan allow NSFW output?
For the open-weight releases the answer is technical rather than legal: the model card states no safety filter, so whoever runs Wan2.2 locally decides what it produces. That is why open checkpoints appear behind services that hosted providers refuse to power.
For Wan2.7 and Wan3.0 the API sits behind Alibaba Cloud, and the release pages we read on Sep 3, 2026 publish no NSFW policy for these endpoints. Absence of a rule is not permission: platform terms and local law apply, and intimate imagery of a real person without consent stays unlawful wherever the model runs.
How is Wan used in tools we review?
None of the six partner tools in our rankings documents its backbone model, so we do not attribute outputs to Wan unless a vendor states it. Structurally, API-based services such as the one in the Wavespeed review expose model catalogs, and open Wan checkpoints are the kind of component such catalogs can list.
In editorial tests the question is not which model a tool uses but what happens to the input image: retention, moderation and deletion. A tool built on an unfiltered open model without its own consent checks scores lower on the privacy axis, weighted 30 percent.
What should a reader check before trusting a Wan clip?
Open weights mean no automatic watermark and no provenance label, so a Wan2.2 clip carries nothing that identifies it as synthetic unless the publisher adds a C2PA manifest or a visible mark. Whoever publishes such clips carries the labeling duty under EU AI Act Article 50.
Keep the prompt, seed and checkpoint name; without them a clip cannot later be shown to be synthetic. See image-to-video for how the input frame constrains what a model can change.
Consent note: tools in our rankings are documented for use only with photos of yourself or consenting adults 18+. See Responsible AI; if an image of you was misused, report NCII.
Sources
- Wan-Video organization on GitHub — repositories Wan2.1 and Wan2.2, accessed Sep 3, 2026.
- Wan2.2 repository and LICENSE.txt — Apache 2.0, uploaded Jul 28, 2025, accessed Sep 3, 2026.
- Wan-AI on Hugging Face — Wan2.2 checkpoints, accessed Sep 3, 2026.
- Alibaba Cloud Model Studio — newly released models — Wan2.7 Apr 3, 2026; Wan3.0 Aug 6, 2026, accessed Sep 3, 2026.
- Alibaba Cloud blog — Wan3.0 at general availability — GA Aug 24, 2026, per-second pricing, accessed Sep 3, 2026.
Reference entry by the Clothoff AI Editorial Team; not legal advice. The access date marks the last verification.
Frequently asked questions about Wan
Is Wan open source?
Partly. Wan2.1 and Wan2.2 are published as open weights, and the Wan2.2 repository carries the Apache 2.0 license, which permits commercial use. Later versions, Wan2.7 and Wan3.0, are closed models that Alibaba Cloud serves only through the Model Studio API, so their weights cannot be downloaded.
What hardware does Wan2.2 need to run locally?
The 5B hybrid model is aimed at consumer graphics cards, while the 14B text-to-video and image-to-video variants need far more video memory or multi-GPU setups. The repository publishes example commands; the exact requirement depends on resolution, clip length and inference framework.
Does Wan add a watermark to generated video?
The open-weight model cards describe no watermark or provenance metadata, so a locally generated clip is unmarked unless the operator adds a C2PA manifest or visible label. The Wan3.0 announcement we read states no watermarking policy for the API either, so provenance stays with the publisher.
Can Wan generate video from a reference image of a person?
Yes, image-to-video and reference-to-video are supported modes in both the open and the API lines. Legally, that capability is exactly where consent matters: animating a real, identifiable person into intimate content without documented permission is unlawful in the United States and elsewhere, whatever the model.
What is the difference between Wan2.7 and Wan3.0?
Wan2.7 appeared in Model Studio on Apr 3, 2026 with separate text-, image- and reference-to-video endpoints. Wan3.0, listed on Aug 6, 2026 and generally available from Aug 24, 2026, accepts text, images, audio, video and documents as reference inputs and extends clip length, per Alibaba Cloud.
Do the tools we review disclose that they use Wan?
No partner in our rankings names its backbone model on the pages we tested, so we do not attribute their output to Wan or any other family. Scores rest on observable behavior, mainly upload handling, retention, moderation and deletion, rather than on which model a vendor licenses.