No, Wan 3.0 is not open source. As of 5 November 2026, Alibaba has published no Wan 3.0 weights, code or model licence, and the model runs only as a hosted service. The newest open Wan video flagship is still Wan 2.2, released under Apache 2.0 in July 2025.

We checked Alibaba's own channels on the date above instead of relying on summaries, because the web is full of pages that say the opposite. This post shows what we looked at, where the "open weights" story came from, and which Wan versions you can really download.

Are the Wan 3.0 weights on Hugging Face or GitHub?

No. These are the results of our check on 5 November 2026.

Where we lookedWhat is thereWan 3.0 found?
Wan-AI on Hugging Face27 model repos, all from the Wan 2.1 and Wan 2.2 lines, plus Wan-Dancer-14BNo
Wan-Video on GitHubSix repos: Wan2.1, Wan2.2, Wan-Animate-2, Wan-Dancer, Wan-skills and a diffusers forkNo
Alibaba Cloud Model Studio docswan3.0-video and wan3.0-video-prime, listed as hosted modelsHosted only, with no mention of weights
ModelScope (Wan-AI)The page did not load its model list for usNot confirmed

We read the full Hugging Face list through its public API, so the count covers every repo and not only those on the first page. None of the 27 names contains Wan3, Wan2.5, Wan2.6 or Wan2.7.

Press coverage agrees. The Next Web's launch report says plainly that the weights are closed, and calls this a shift from Alibaba's history of open-sourcing the Wan series. It names Wan 2.2 as the last video flagship Alibaba released as open source.

We found no Alibaba statement that promises Wan 3.0 weights, and none that rules them out. Treat the question as "not released, no date given", not as "never".

What is the difference between open source and open weights?

The two terms get mixed up a lot, and it matters here.

  • Open weights means you can download the trained model files and run them on your own hardware.
  • Open source usually goes further: the inference code and a licence that lets you use, change and share the work.
  • Closed or hosted means you send a request to someone's server and get a video back. You never hold the model.

Wan 2.2 shipped inference code and weights together under Apache 2.0. Wan 3.0 fits the third group. You can use it, but only through a service that hosts it.

When was Wan 3.0 released?

Wan 3.0 entered public beta in early August 2026, with 6 August the date most often given, and Alibaba opened it more widely on 24 August 2026. The Next Web reported the wider launch that same day. So if you searched for a Wan 3.0 release date, the model is out. Only the weights are missing.

The hosted version is not a small preview. Alibaba's Model Studio docs list text-to-video, first and last frame input, all-in-one references, audio, 480p to 1080p and clips of 2 to 30 seconds. Launch pricing was $0.05, $0.10 and $0.20 per second at those three resolutions, according to The Next Web.

Why do some sites say Wan 3.0 is open source?

Because the claim got copied. Several unofficial sites describe an Apache 2.0 release of Wan 3.0 with 1.3B and 14B checkpoints, some with dates as early as April 2026. That date is months before the public beta began, and none of those checkpoints exists in the official Hugging Face or GitHub organisations.

The confusion has a likely source. Alibaba has kept publishing open Wan work in 2026, just not Wan 3.0:

  • Wan2.2-Animate-2-14B, a character animation model, is on Hugging Face. The Wan-Animate-2 repo on GitHub carries an Apache 2.0 licence.
  • Wan-Dancer-14B appeared on Hugging Face in July 2026.

Both are real, recent and from the official organisations. Neither is Wan 3.0. A new open Wan release in the same month as the Wan 3.0 beta is easy to misread, and the Wan-Animate-2 repo itself does not mention Wan 3.0.

Which Wan versions are open, and which are closed?

Here is where each version stands as of 5 November 2026.

VersionPublic weightsLicence for the weightsHow you run it
Wan 2.1Yes (February 2025)Apache 2.0Self-host or hosted
Wan 2.2Yes (28 July 2025)Apache 2.0Self-host or hosted
Wan 2.5NoNone publishedHosted only
Wan 2.6NoNone publishedHosted only
Wan 2.7NoNone publishedHosted only
Wan 3.0NoNone publishedHosted only

The split is clear: the open line stops at Wan 2.2. Every numbered release after it has been hosted only.

Is Wan 2.5 open source?

No. Wan 2.5 is not open source either. No Wan 2.5 repo exists in the official Hugging Face or GitHub organisations, and Alibaba's docs list it as wan2.5-t2v-preview and wan2.5-i2v-preview, both hosted. If a site offers a Wan 2.5 checkpoint, it did not come from these official organisations.

What does Apache 2.0 let you do with Wan 2.2?

Apache 2.0 is a permissive licence that allows commercial use, changes and redistribution, subject to its terms. The Wan 2.2 repo adds that the team claims no rights over your generated content, and that you are responsible for how you use it. Illegal content, harm to people, misinformation and targeting vulnerable groups are all ruled out.

Wan 3.0 has no such licence for you to rely on. Your rights come from the terms of whichever service you use, so read those before you ship commercial work.

Can you run Wan 3.0 locally?

No, because there is nothing to download. A local Wan 3.0 install, a ComfyUI workflow labelled Wan 3.0 or a "Wan 3.0 download" page cannot be running the official model, since the official files are not public. Do not run installers from those pages.

If you need a Wan model on your own machine, Wan 2.2 is the answer. The repo gives these single-GPU figures:

Wan 2.2 modelTaskOutputStated single-GPU need
TI2V-5BText and image to video720pAt least 24GB VRAM, such as an RTX 4090
T2V-A14BText to video480p, 720pAt least 80GB VRAM
I2V-A14BImage to video480p, 720pAt least 80GB VRAM

Our guide on how to use Wan 2.2 covers both the online and the local route.

How can you tell if Wan 3.0 weights are really out?

Rumours will keep appearing, so use this four-step check when you see a new claim.

  1. Open the official Hugging Face organisation. Look for a Wan 3 repo under Wan-AI, not a look-alike account with a similar name.
  2. Open the official GitHub organisation. A real release should come with a repo under Wan-Video that holds inference code and a licence file.
  3. Read the licence on the model card. Do not trust a licence quoted on a third-party page.
  4. Match the model name. Wan2.2-Animate-2 and Wan-Dancer are real, but they are not Wan 3.0.

If those steps fail, the claim is not confirmed, however confident the page sounds.

How do you use Wan 3.0 without the weights?

You use a hosted service. That means Alibaba Cloud Model Studio, a third-party API, or a studio that already runs the model. On Kenerate, Wan 3.0 runs in the video studio with no Alibaba Cloud account, and you pay per clip in credits.

These are the current Kenerate rates, in credits per second, from our model configuration.

Model480p720p1080p
Wan 3.0153060
Wan 3 Prime22.54590
Wan 2.7not offered3045
Wan 2.6not offered3045
Wan 2.5153045
Wan 2.2918not offered

A 5-second Wan 3.0 clip at 720p costs 150 credits. Wan 3.0 on Kenerate takes text, a start image or up to 10 reference images, and supports lengths of 2 to 15, 20, 25 or 30 seconds. New accounts get 25 credits, which does not cover the shortest Wan 3.0 clip (30 credits for 2 seconds at 480p). Our post on whether Wan 3.0 is free goes through the free routes in detail.

Should you wait for open weights or use a hosted model?

It depends on what you need from the model.

  • Use Wan 2.2 on your own GPU if you need local control, fine-tuning or a clear licence.
  • Use Wan 3.0 hosted if you need longer clips, sound or references, and you are fine with paying per second.
  • Waiting for weights is a poor plan on its own, because no date has been given.

If you are still choosing a version, compare them in Wan 2.2 vs Wan 2.7 vs Wan 3, or see the whole family on the Wan model page.