Short answer: no official open-weight release of Wan 3.0 has been verified. Alibaba’s official Wan-Video organization on GitHub lists no Wan 3.0 repository, and the official Wan-AI organization on Hugging Face lists no Wan 3.0 checkpoint. Wan 3.0 is documented in Alibaba Cloud Model Studio under the model ID wan3.0-video and is currently marked as preview.
If you found a page offering a “Wan 3.0 open source download,” an Apache 2.0 checkpoint, or a ComfyUI node that loads Wan 3.0 weights, that page describes something Alibaba has not published.
This guide covers what “open source” actually means for Wan 3.0, what Alibaba genuinely released, why the claims online conflict, which Wan model you can download and run today, and how to check the status yourself rather than relying on any single article — including this one.
The Short Answer: Wan 3.0 Ships Through Hosted Access
The most accurate statement is narrow and checkable:
Wan 3.0 does not currently have a verified official open-weight release.
Three signals support that.
The official GitHub organization has no Wan 3.0 repository. Alibaba’s Wan-Video organization currently hosts six public repositories, covering Wan 2.1, Wan 2.2, Wan-Dancer, Wan-Animate-2, an agent-skills project, and a Diffusers fork. No Wan 3.0 model repository is listed.
The official Hugging Face organization has no Wan 3.0 checkpoint. A search of the Wan-AI organization for Wan 3.0 models returns zero results. Its published models cover the Wan 2.2 family, Wan-Dancer, and Wan 2.2 Animate variants.
Alibaba documents Wan 3.0 as a cloud API, not a download. The official Wan 3.0 API reference describes an asynchronous HTTP service with Beijing and Singapore endpoints, an API key requirement, and the model marked as preview. Nothing in that documentation points to downloadable files.
So when people ask whether Wan 3.0 is open source, they are usually asking one of these:
Can I download official Wan 3.0 weights?
Can I clone an official Wan 3.0 repository?
Can I self-host Wan 3.0 from an official checkpoint?
Can I run Wan 3.0 locally in ComfyUI?
For all four, no verified official release currently supports that workflow.
Open Source, Open Weights, and Hosted Access Are Three Different Things
Much of the confusion comes from collapsing three kinds of availability into one phrase.
Open-source code is published under a license defining how it can be inspected, modified, and redistributed. A public repository alone is not enough — the license is what matters. An API wrapper, SDK, or interface can be open source while the underlying model is not.
Open model weights are the trained files needed to run inference yourself. For anyone doing local video generation, this is the part that counts. Downloadable weights let you run on private hardware, deploy on your own infrastructure, experiment with quantization, build ComfyUI or Diffusers workflows, and fine-tune where the license permits.
Hosted access means the model runs on someone else’s infrastructure. You send a prompt and reference assets, inference happens remotely, and a video comes back. This is what Alibaba documents for wan3.0-video.
Based on Alibaba’s currently published materials, Wan 3.0 is available through hosted access, while no official open-weight release has been verified. It follows that free access is not the same thing either — several platforms let you generate with Wan 3.0 without paying, but that is free hosted access, not an open-weight release.
Measured against what an open release would normally involve, Wan 3.0 currently has no published model weights, no public inference code, no stated model license, no official repository, and no local ComfyUI workflow. This is not a case of open weights under restrictive terms, or code released with weights pending. What Alibaba has opened for Wan 3.0 is access.
Is Wan 3.0 on GitHub?
No official Wan 3.0 model repository is currently listed under Alibaba’s Wan-Video organization.
You will still find “Wan 3.0” in GitHub search results. Third-party developers publish API clients, workflow templates, automation scripts, prompt collections, web interfaces, and ComfyUI API integrations. Those projects may work perfectly well against a hosted endpoint, but none of them contains the model.
When evaluating any supposed Wan 3.0 GitHub download, ask three questions:
Is the repository published under the official Wan-Video organization?
Does it contain, or officially link to, actual Wan 3.0 model weights?
Is there a model license defining how those weights may be used?
If any answer is no, the repository is not evidence of an open Wan 3.0 release.
Is Wan 3.0 on Hugging Face?
The place to check is the official Wan-AI organization. Searching it for Wan 3.0 models returns nothing.
This matters because Hugging Face also hosts community uploads, discussions, third-party quantizations, Spaces, and unofficial model cards. A page with “Wan 3.0” in its title is not automatically an Alibaba release.
A genuine open-weight release would be straightforward to verify, because you would expect several signals together:
A model card under the Wan-AI organization
Downloadable model files
Architecture or parameter information
A stated license
Inference instructions
An official GitHub cross-link
Those signals are all present for earlier open Wan releases. None is present for Wan 3.0.
Why Wan 3.0 Open-Source Claims Conflict Online
If you landed on a confident page describing Apache 2.0 weights with 1.3B and 14B checkpoints, you are not alone. Several high-ranking pages say something along those lines. None of it has primary-source support in Alibaba’s published materials, and the reason those pages exist is worth understanding.
The Wan line has a real open-release history. Wan 2.1 and Wan 2.2 both shipped with downloadable weights and inference code, so the community reasonably expects each new version to follow the same pattern. When a version number appears, search demand for open weights spikes immediately, well before the release status is settled.
Pages published into that gap tend to carry forward assumptions from earlier Wan releases without official Wan 3.0 confirmation. Parameter counts get borrowed from previous generations. Licence names get assumed. Installation instructions get adapted from Wan 2.2 documentation. The same pattern produced conflicting Wan 2.7 pages before Wan 3.0 launched, and it produces other unsupported Wan 3.0 details today: native 4K output, a 60B mixture-of-experts architecture, twelve reference image slots, a six-shot director mode. None of these appears in Alibaba’s documentation.
The rule that resolves it: do not infer Wan 3.0’s licence or download status from Wan 2.2. Verify Wan 3.0 independently against official channels.
What Alibaba Actually Released With Wan 3.0
Setting the open source question aside, the launch was substantial. These parameters come from the official Wan 3.0 API reference rather than secondhand spec lists.
Model ID and status. The model is wan3.0-video, described as an all-in-one reference-based model supporting text-to-video, image-to-video with first-frame and first-and-last-frame control, and reference-based generation. The API reference marks it preview, and access is granted on request rather than open to every account.
Duration: 2 to 30 seconds. Without video input, the duration parameter accepts an integer in that range, defaulting to 5. With video input, input plus output must not exceed 30 seconds total. Setting duration to -1 activates smart duration mode, where the model recommends a length based on the prompt and supplied media.
Resolution: 480P, 720P, 1080P — default 1080P. Those are the only documented tiers. There is no 4K option.
Six aspect ratio values, with adaptive as the default. The ratio parameter accepts adaptive, 16:9, 4:3, 1:1, 3:4, and 9:16. Adaptive automatically recommends a ratio based on the proportions of the input media and the generation intent.
Audio on by default. The audio parameter is a boolean defaulting to true, so a Wan 3.0 clip comes back with sound unless you explicitly disable it.
Reference limits. A request can include up to 10 reference images, 5 reference video clips totalling no more than 15 seconds, and 5 reference audio clips totalling no more than 15 seconds. Reference inputs and first-frame/last-frame inputs are mutually exclusive and cannot be combined in one request.
Document and webpage input. Wan 3.0 accepts one file or one web link per request, never both. Supported file formats are docx, doc, xlsx, xls, pptx, ppt, pdf, txt, key, pages, numbers, and md, up to 100MB, with a 50-page limit validated for the paginated formats. Web links must point to publicly accessible pages that do not require a login.
For a fuller breakdown of the launch and how the specs compare against earlier versions, see our Wan 3.0 release date guide.
Wan 2.2 Is the Wan Model You Can Actually Download
If your real goal was local deployment, the useful answer is not to wait for Wan 3.0 weights. It is Wan 2.2, which is the most recent general-purpose Wan generation with clearly verified official downloadable weights.
Alibaba released the Wan 2.2 inference code and model weights on July 28, 2025 under Apache 2.0, on both Hugging Face and GitHub, in three main variants: T2V-A14B for text-to-video, I2V-A14B for image-to-video, and TI2V-5B, a smaller model handling both tasks at 720P and 24fps. Per its official model card, TI2V-5B runs on a single consumer GPU, which makes it the practical entry point for most local setups. Wan 2.2 was also integrated into ComfyUI and Diffusers on release day, so the local tooling is mature rather than experimental.
One clarification worth making, because it is easy to get wrong in the other direction: the Wan team has not stopped releasing open models. Wan-Dancer and Wan-Animate-2 are both Apache 2.0 and both newer than Wan 2.2. What has not been released openly is any model from the 3.0 generation, and Wan 2.2 remains the most recent open general-purpose video foundation model in the line.
Does Wan 3.0 Work With ComfyUI?
This needs the same distinction as everything else on this page.
There is no verified official Wan 3.0 checkpoint that a local-weight ComfyUI workflow can load, because no Wan 3.0 model file has been published.
A third-party developer could still build a ComfyUI custom node that sends requests to a hosted Wan 3.0 API. That is a legitimate thing to build, but it would not make Wan 3.0 open source and it would not mean the model runs on your machine.
So when you see “Wan 3.0 ComfyUI,” check which of these it is:
A. A workflow that loads Wan 3.0 model weights locally — not currently possible. B. A node that calls a remote Wan 3.0 API — possible, but that is hosted access wearing a local interface.
Wan 2.2, by contrast, has had genuine local ComfyUI support since its release day.
How to Verify Wan 3.0 Open-Source Status Yourself
Release status changes. Rather than relying on any article, check primary sources. Four signals together are far more reliable than any one alone.
Check the Wan-Video organization on GitHub. Look for a Wan 3.0 model repository under the official organization, not a similarly named third-party repo. Check whether it actually contains model code, official weight links, and licensing information.
Check the Wan-AI organization on Hugging Face. A real release appears as an official model card with downloadable files, stated parameter size, supported tasks, and hardware requirements. Verify the publisher, because community uploads often use similar names.
Read the licence. Downloadable does not mean open source. Check what is actually permitted for commercial use, redistribution, modification, hosted services, and derivative models. Do not assume Wan 2.2’s Apache 2.0 terms carry over.
Look for official cross-linking. The strongest signal is Alibaba’s GitHub, Hugging Face organization, ModelScope, Wan website, and Model Studio documentation all pointing at the same release. If a checkpoint exists but no official channel references it, treat it with caution.
“Not Open Source Yet” Is Not “Closed Forever”
A lot of coverage gets this wrong in both directions.
The evidence supports: no verified official open-weight Wan 3.0 release currently exists.
The evidence does not support: Alibaba will never release Wan 3.0 openly.
The Wan team has released open models before and continues to. Alibaba could later publish a full Wan 3.0 checkpoint, a smaller variant, selected task-specific models, distilled versions, inference code, or an open release following an initial hosted period.
All of those are speculative until announced. Claims that Wan 3.0 is already Apache 2.0, and claims that it will never be opened, both go beyond what the sources currently show. If the status changes, the official Wan-AI organization is where it will appear first.
Local Wan 2.2 vs Hosted Wan 3.0: How to Choose
Consideration | Wan 2.2 (local) | Wan 3.0 (hosted) |
|---|---|---|
Cost model | Hardware and electricity | Per generated second |
Maximum clip length | Short clips, extended by stitching | 30 seconds in one pass |
Audio | Not generated with the video | On by default, same pass |
Document input | No | 12 file formats, plus web links |
Fine-tuning and LoRA | Yes | No |
Offline and private | Yes | No |
Setup effort | Substantial | None |
Availability | Permanent, you hold the weights | Subject to platform access |
The decision comes down to one question: do you need control, or do you need capability? Choose Wan 2.2 if you need to fine-tune, work offline, keep material private, or generate at volume without per-second costs. Choose Wan 3.0 if you need 30-second continuous takes, synchronized audio, or document input — none of which any downloadable Wan model can currently do.
Many teams use both: Wan 2.2 locally for iteration, hosted Wan 3.0 for the small number of finished pieces that need the longer form and the audio.
How to Use Wan 3.0 Without Downloading It
Because no official downloadable Wan 3.0 checkpoint is currently available, most users access the model through a hosted service. That removes the need to fetch large checkpoints, configure CUDA dependencies, manage GPU memory, or maintain an inference environment.
You can generate with Wan 3.0 through the Wan 3.0 AI video generator, which exposes the model’s 480p, 720p, and 1080p output, its duration range up to 30 seconds, and its aspect ratio options in a single interface, with no Alibaba Cloud setup and no API key required.
If you would rather work with a model whose behavior is better established while Wan 3.0 preview capacity settles, Wan 2.7 remains the line’s proven workhorse for shorter clips.
Want to use Wan 3.0 without managing local model files?
Frequently Asked Questions
Is Wan 3.0 open source?
No official open-weight release has been verified. Alibaba’s Wan-Video organization on GitHub lists no Wan 3.0 model repository, and the Wan-AI organization on Hugging Face lists no Wan 3.0 checkpoint. Wan 3.0 is documented as a hosted API model.
Can I download Wan 3.0 weights?
Not currently. No Wan 3.0 checkpoint is published through Alibaba’s primary GitHub or Hugging Face organizations. Treat any “Wan 3.0 weights download” or “Wan 3.0 14B” offer as unverified unless it traces back to an official channel.
Is Wan 3.0 on GitHub?
Third-party repositories reference Wan 3.0, but the official Wan-Video organization does not list a Wan 3.0 model repository. Those third-party projects are typically API clients or workflow tooling, not the model.
Is Wan 3.0 on Hugging Face?
The official Wan-AI organization does not list a Wan 3.0 model checkpoint. Community uploads and discussions mentioning Wan 3.0 are not official releases.
Can I run Wan 3.0 locally?
Not from an official checkpoint. Every documented route to Wan 3.0 goes through Alibaba’s API or a platform calling it. Third-party API wrappers should not be mistaken for local inference.
Does Wan 3.0 work in ComfyUI?
Not as a local-weight workflow, because no Wan 3.0 model file has been published. A custom node could call a hosted Wan 3.0 API, but that is remote generation rather than running the model yourself. Wan 2.2 has genuine local ComfyUI support.
Which Wan model can I download?
Wan 2.2, released July 28, 2025 under Apache 2.0, in T2V-A14B, I2V-A14B, and TI2V-5B variants, with weights and inference code on Hugging Face and GitHub.
Does Wan 3.0 have an Apache 2.0 licence?
No Wan 3.0 model licence has been published. Wan 2.2’s Apache 2.0 terms should not be assumed to apply to Wan 3.0.
Does Wan 3.0 support 4K?
No. The official API reference documents 480P, 720P, and 1080P only, with 1080P as the default. The 4K claim appears in many third-party comparison tables but has no official source.
Will Wan 3.0 become open source later?
No release has been announced. Because earlier Wan generations were released openly, a future release is possible, but it should not be treated as confirmed until it appears on the official Wan-AI organization.
The Bottom Line
Wan 3.0 has not received a verified official open-weight release. What Alibaba shipped is a capable hosted model: 2-to-30-second generation with smart duration, 480P to 1080P output, adaptive aspect ratio, audio on by default, up to 20 reference assets, and document or webpage input — none of which you can currently run on your own hardware.
If you need weights, Wan 2.2 is the real answer, and a genuinely strong one. If you need what Wan 3.0 does, the question is which access route suits you, not which repository to clone.
Want to use the model now?