The Newest Wan You Can Download Is Still Wan 2.2

Introduction

Wan 2.2 shipped on July 28, 2025 under Apache 2.0. Fourteen months later it is still the newest Wan that anyone outside Alibaba can download.

In between, Alibaba has shipped four more releases under the same brand: Wan 2.5-Preview in September 2025, Wan 2.6 announced December 16, 2025, the four-model Wan 2.7 suite in April 2026, and Wan 3.0 in public beta in early August 2026 with Alibaba's own write-up following on August 13. None of the four published weights. The line that built its reputation as the open-weights answer to Veo and Sora has been closed for four consecutive generations while keeping the name that earned the reputation.

That gap is worth writing about on its own. What makes it worth writing about carefully is that the search results describing it are actively wrong — several sites state as fact that Wan 2.7 is Apache 2.0 with weights on Hugging Face. It is not, and you can prove that yourself in about thirty seconds. This post shows you how, because the method outlives this particular model.

Four releases, no weights

Alibaba has never announced a policy change here, and that is precisely how the confusion survives. The announcements are consistent by omission: the Wan2.7-Video post says the models are "now available on Alibaba Cloud's Model Studio and the official Wan website," and the Wan3.0 post says it is "now available on Alibaba Cloud Model Studio." Neither sentence mentions open source, a repository or a license. Alibaba Cloud's Model Studio API reference for Wan3.0, last updated September 4, 2026, contains zero occurrences of the strings "Apache," "weights," "license" or "open source" — it is an endpoint document, and it reads like one.

A reader who remembers Wan as the open model, and who sees a launch post that simply does not raise the subject, will fill the silence with the prior. That is the whole failure mode.

Verify it yourself in thirty seconds

Two HTTP requests settle it. Neither needs an account.

Hugging Face exposes a public model search. Ask it for anything matching the name:

curl -s 'https://huggingface.co/api/models?search=Wan2.7'

Run today, that returns exactly this:

[]

An empty array, hub-wide — not "Alibaba hasn't uploaded it," but "no repository on Hugging Face, from any account, has that string in its name." The same query for Wan3.0 returns the same empty array. And listing the vendor's own organization shows what does exist:

curl -s 'https://huggingface.co/api/models?author=Wan-AI&limit=100'

Twenty-seven repositories. Every one is a Wan2.1 or Wan2.2 variant, a -Diffusers conversion of one, or one of the two auxiliary models discussed below. Nothing numbered above 2.2.

GitHub answers the second half:

curl -s 'https://api.github.com/orgs/Wan-Video/repos?per_page=100'

Six repositories: Wan2.1, Wan2.2, Wan-skills, Wan-Dancer, Wan-Animate-2, and a fork of diffusers. No Wan2.7. No Wan3. The inference code that would accompany an open release does not exist either.

Both endpoints return JSON, both are unauthenticated, and both are far stronger evidence than the thing people usually rely on — the absence of a download button on a page they happened to read. Absence of a link is ambiguous. An empty array from the canonical index is not.

Generalise the pattern: https://huggingface.co/api/models?search=<name> plus https://api.github.com/orgs/<org>/repos is a thirty-second check you can run against any weights claim, including the ones on this site.

The SEO layer says the opposite, and contradicts itself

Search for "wan 2.7 open source" and the results are confident and wrong. The clearest example is wan27.org, an affiliate site built around the Wan brand, whose "Wan 2.7 Open Source Guide" opens with:

Is Wan 2.7 open source? Yes — open-weight under Apache 2.0. See what is actually open, where to download weights, hardware requirements, and local setup with ComfyUI or Python.

It then explains what Apache 2.0 permits, promises a section on "Where to Get the Weights: HuggingFace, GitHub, and ModelScope," and quotes disk requirements for a download that does not exist.

The striking part is that the same domain also publishes a Wan 2.7 Hugging Face guide, which says the opposite. Of the Wan-AI organization, that page reports:

...it has not been updated with Wan 2.7 weights. Here is what the Wan-AI org currently hosts: Wan2.2-T2V-14B, Wan2.2-I2V-14B, Wan2.1-T2V-14B. These are solid models, but if you specifically want Wan 2.7, HuggingFace alone will not get you there.

Both pages are live. Both are indexed. One tells you the license terms of a checkpoint the other tells you does not exist. This is what happens when pages are generated per keyword rather than per fact: "wan 2.7 open source" gets an article that answers yes because that is the shape of the query, and "wan 2.7 huggingface" gets an article that answers no because that query resolves against a real page someone actually looked at.

I could not verify a similar claim on cliprise, which is currently a parked domain listed for sale rather than a live site. Treat any specific attribution beyond wan27.org as unchecked; the pattern, not the roster, is the point.

What is actually behind the paywall

The closed releases are not cosmetic, which is why the confusion costs people real time. Compared with what you can download:

• Audio. wan3.0-video generates sound alongside the picture by default. The open Wan 2.2 text-to-video and image-to-video models output silent video; Wan2.2-S2V-14B consumes an audio track rather than synthesising one.
• Length and resolution. Wan 3.0 runs 2–30 seconds at 480P, 720P or 1080P, 30fps. The open models do five seconds at 480P/720P.
• Document and webpage input. Model Studio's Wan3.0 reference accepts one file up to 100MB — pptx, xls, pdf, txt, md — or one public URL, and builds video from it. No open Wan does this at all.
• Voice-carrying references. wan2.7-r2v takes a reference_voice audio URL attached to a reference image or video, carrying a speaker's timbre into new footage.

Note also the asymmetries inside the closed tier, because the headline numbers hide them: wan2.7-videoedit has no audio and caps at 10 seconds, and wan2.7-r2v also caps at 10 — only t2v and i2v reach 15. And thinking_mode, which the marketing attaches to the 2.7 generation generally, is documented only for wan2.7-image and wan2.7-image-pro; the Wan 3.0 video API exposes prompt_extend instead.

The hard constraint for local users is simpler than any of that. There is no consumer-hardware path above Wan 2.2. No 5B Wan 3.0, no quantized Wan 2.7, no RTX 4090 route to either. The 24GB story ends at Wan2.2-TI2V-5B, a July 2025 model, and community GGUF builds of the A14B pair — which is why that generation still moves millions of downloads a month more than a year on.

Alibaba did not exit open weights, it moved them off the flagship

The easy version of this story — "Alibaba closed Wan" — is wrong, and the accurate version is more interesting.

Tongyi Lab is still shipping Apache 2.0 video weights. Wan-Dancer-14B, a music-to-dance model, went up in July 2026. Wan-Animate-2, for character animation from a driving video, followed in August 2026 with both weights and inference code. Both are Apache 2.0 on the Hugging Face model cards; both were published after Wan 2.7 shipped closed, and one of them landed the same month as Wan 3.0.

And both are built on the Wan 2.2 base.

That is the actual policy, inferred from behaviour rather than announced: the open branch continues, but it is derivative and task-specific, downstream of a base model that stopped advancing fourteen months ago. Frontier text-to-video is a product. Character animation, dance and LoRA fine-tuning infrastructure are ecosystem. The Mixture-of-Experts video generation model at the top of the stack is not coming back to Hugging Face on current evidence, and two GitHub issues asking whether 2.5 and up would be opened have sat unanswered since September 2025.

The same split, three other labs, three different asterisks

This is not a Wan story so much as the shape of open weights in late 2026 — and the asterisk is never in the same place twice.

Qwen3.8-Max opened its 2.4-trillion-parameter flagship on August 12, 2026, and the headline was accurate as far as it went: the weights are real and downloadable. The license is a custom revenue-gated document rather than Apache 2.0, and the open checkpoint is text-only with a 262,144-token native window against the hosted model's 1M — the caveats live in the artifact, not the availability.

Tencent's Hy4 preview is the cleanest of the three: 770B MoE, genuine Apache 2.0, published August 28, 2026 — with "preview" doing the qualifying work instead of the license.

MiniMax H3 shipped weights for a frontier video model, which is rarer than it sounds, under a community license that bars local deployment in the US, EU, UK and South Korea without written authorization. Open in the repository, closed in the jurisdictions where most of this site's readers work.

Wan's asterisk is the one that is hardest to see, because it is not in a license at all. It is in a version number. Nothing about Wan 3.0 is presented as restricted; the restriction is that the artifact was never published, and the brand keeps vouching for a commitment that expired four releases ago.

Conclusion

Wan 2.2 remains an excellent open model and the best local video generation available to a 24GB GPU. It is also fourteen months old, and every Wan release since — including the current flagship, which independent arena voting places among the strongest video models in the field — exists only as a metered API endpoint.

If you are planning around Wan, plan around that. Assume 2.2 is the last open frontier release, because for four generations it has been. If you need audio, thirty-second clips or 1080P, you need Alibaba Cloud and a per-second bill. And when a page tells you a model's weights are open, spend the thirty seconds: curl the Hugging Face search API and list the vendor's GitHub org. An empty array is the whole answer, and it does not care how confident the article was.

Sources

• Alibaba Unveils Wan2.7-Video — Alibaba Cloud Community, April 7, 2026. The four-model suite, 2–15s, 720p/1080p; no license or weights mentioned
• Wan3.0: 30-Second AI Video Generation from Any Input — Alibaba Cloud, August 13, 2026
• Wan3.0 - Video Generation API Reference — Alibaba Cloud Model Studio, last updated September 4, 2026. wan3.0-video and wan3.0-video-prime, 30s at 30fps, prompt_extend, "currently in preview"
• https://huggingface.co/api/models?search=Wan2.7 — returns []
• Wan-AI on Hugging Face — 27 model repositories, all Wan 2.1, Wan 2.2, Wan-Animate-2 or Wan-Dancer-14B
• Wan-Video on GitHub — six repositories; no Wan2.7, no Wan3
• Wan-Video/Wan2.2 — the last open frontier release, Apache 2.0, July 28, 2025
• Wan-AI/Wan-Dancer-14B and Wan-AI/Wan2.2-Animate-2-14B — Apache 2.0 on the model cards, July and August 2026
• Our catalog pages: Wan 3.0, MiniMax H3, Tencent Hy4 preview, Qwen3.8-Max, and the ComfyUI and Wan 2.1 tool pages