Wan 3.0: What Is Actually Confirmed
Alibaba's Wan 3.0 went live on QwenCloud as an application-gated test on August 6, 2026 — it is not open to everyone, and its weights have not been published.
Most of the specs circulating for Wan 3.0 have no official source behind them. This page separates what Alibaba and QwenCloud have actually documented from what is only being repeated, names the source for every line, and logs the date each source was checked. Last verified August 7, 2026.
Wan 3.0 is not available in our generator
You cannot generate with Wan 3.0 on this site. No Wan model of any version is available here. This page exists to document the model's status accurately, not to sell access to it — access is granted by Alibaba, through QwenCloud, on application.
All sources last checked on
Official video
Alibaba's own Wan 3.0 announcement footage
The launch video Alibaba published for Wan 3.0 on August 6, 2026. Every shot in it was rendered by Alibaba — none of it was generated on ImagineToVideo, and none of it has been independently reproduced. Treat it as vendor marketing, not as a benchmark.
What QwenCloud's own documentation states
2 to 30 seconds per request
The API reference lists a duration range of 2–30 seconds, plus a smart-duration setting that lets the model pick. When a source video is supplied, the combined result still caps at 30 seconds.
One model, several input modes
The release log describes a single unified model covering text-to-video, image-to-video from a first frame or a first-and-last frame pair, and reference-to-video — rather than separate checkpoints per mode.
Seven reference input types
The documented reference fields are first frame, last frame, reference image, reference video, reference audio, file and link — which is what the announcement's "Omni-Reference" wording refers to.
Audio generated by default
The audio parameter defaults to on, and a watermark toggle defaults to off. Output resolution is documented as 480P, 720P or 1080P, with 1080P as the default.
Confirmed, with the source for each line
Every statement below comes from Alibaba or QwenCloud directly. Nothing here is sourced from an aggregator or a third-party spec page.
Wan 3.0 became available on QwenCloud on August 6, 2026, as an invite-only test.
QwenCloud's model release log carries the entry on that date and describes it as invite-only testing. Alibaba's own announcement calls it a Public Beta you apply for. Both descriptions mean the same thing in practice: access is granted on request, not open to everyone.
The documented model identifier is wan3.0-video.
A create-task API reference page exists under that identifier, which is why the parameters below can be quoted rather than inferred.
Clip length runs from 2 to 30 seconds, with a smart-duration option.
The duration parameter accepts an integer in that range. Supplying a source video does not extend the ceiling — the combined output still stops at 30 seconds.
The documented resolution options are 480P, 720P and 1080P, defaulting to 1080P.
This matters because it is the single most misreported figure about this model. See the claims table below for what that does and does not settle.
Seven reference input fields are documented, including files and links.
First frame, last frame, reference image, reference video, reference audio, file and link. Aspect ratio can be set to adaptive or to one of five fixed ratios, and the prompt field accepts up to 20,000 characters.
No pricing has been published for Wan 3.0.
Neither the release log entry nor the API reference states a rate. Any per-second or per-clip figure you have seen quoted did not come from Alibaba.
Claims in circulation, checked one by one
These are the Wan 3.0 specs being repeated across the web. For each one, what the official sources actually say.
| Claim in circulation | What the official sources say | Verdict |
|---|---|---|
Native 4K output | Contradicted by the documented parameter list This settles what the API accepts today. It is not a statement about what the model may support later. | |
Released under Apache 2.0 as open weights | ⚠️ Unconfirmed — nothing published so far Absence to date. It does not rule out a later release. | |
27B or 60B mixture-of-experts architecture | ⚠️ Unconfirmed — no official figure exists | |
An "Identity Lock" character-consistency feature | ⚠️ Unconfirmed — name not used in any official source | |
An "AI Director" shot-planning mode | ⚠️ Unconfirmed — name not used in any official source | |
Open to the public as a beta anyone can join | Partly right — application-gated, not open access The two official wordings differ; both require an application. |
Every verdict above reflects the state of the official sources on 2026-08-07 and is re-checked weekly.
The sources, and what each one showed
The four places a Wan 3.0 claim can be settled, and what was found in each on the verification date.
- Checked
- 2026-08-07
- What it showed
- Entry dated 2026-08-06 for wan3.0-video, described as invite-only testing. No resolution, duration, licence, parameter count or pricing stated.
- Checked
- 2026-08-07
- What it showed
- Live, with a full parameter list: 2–30s duration, 480P/720P/1080P resolution, six aspect ratios, seven reference inputs, audio on by default. Marked invite-only. No pricing.
- Address
- Checked
- 2026-08-07
- What it showed
- No Wan 3.0 weights. The most recent addition to the organisation is Wan-Dancer-14B, published 2026-07-10.
- Checked
- 2026-08-07
- What it showed
- No Wan 3.0 repository. The most recently updated repository is Wan-Dancer, last touched 2026-07-17.
These four are re-checked weekly. Anything that changes is written into the update log at the bottom of this page on the same day.
What this actually means right now
If you make videos
You cannot use Wan 3.0 today unless your application for the test is granted. Nothing about the model's quality has been independently verified — every clip in circulation is either Alibaba's own marketing footage or comes from someone with early access. Plan around tools you can run today, and revisit this page when the status changes.
If you build on video APIs
The request schema is documented, so you can design against it, but access is granted case by case and no rate has been published. Two things worth watching before committing: whether the resolution list grows beyond 1080P, and whether weights ever appear on Hugging Face or GitHub, since that determines whether self-hosting is an option at all.
How to check any Wan 3.0 claim yourself
Four steps, no special access required. If a spec cannot survive these, it is not a spec.
Start at the release log
Open QwenCloud's model release log and find the wan3.0-video entry. The date on it is the release date; anything earlier that you have read is a preview or a rumour.
Read the request schema
Open the wan3.0-video create-task reference. Accepted parameter values are the strongest evidence available for what the model does: a resolution or duration that is not in the list is not currently offered.
Check both weight hosts
Look at the Wan-AI organisation on Hugging Face and the Wan-Video organisation on GitHub. If neither has a Wan 3.0 repository, no open-weights or licence claim can be true yet, whatever a summary page says.
Trace the number to a source
For any remaining figure, follow it back until you reach Alibaba or QwenCloud. Most spec claims about this model trace back to another summary page rather than to a primary source, and stop there.
Update log
The four official sources are re-checked weekly. Every change lands here with its date.
Page published after a full check of all four sources
The API reference was found to be live with a complete parameter list, which moved the 4K claim from unverified to contradicted. Open weights remain absent from both Hugging Face and GitHub. No pricing published.
wan3.0-video appears in QwenCloud's release log
Listed as an invite-only test covering text-to-video, image-to-video from a first frame or a first-and-last frame pair, and reference-to-video in one unified model.
Questions people are asking about Wan 3.0
When was Wan 3.0 released?
QwenCloud's model release log carries the wan3.0-video entry dated August 6, 2026, describing it as an invite-only test. That is the earliest date backed by an official source. Anything describing a release before then is referring to a preview, an ecosystem model such as Wan-Dancer, or speculation.
Can I use Wan 3.0 right now?
Only if your application is accepted. QwenCloud's release log describes invite-only testing and Alibaba's announcement asks readers to apply, so access is granted case by case rather than open to everyone. You cannot use Wan 3.0 on ImagineToVideo — it is not in our generator.
Is Wan 3.0 open source?
No weights have been published as of August 7, 2026. There is no Wan 3.0 repository on the Wan-AI organisation on Hugging Face or the Wan-Video organisation on GitHub. Earlier Wan models were released openly, which is probably why the Apache 2.0 claim spread, but it has not happened for this version. We re-check both hosts weekly.
Does Wan 3.0 support 4K?
Not through the documented API. QwenCloud's create-task reference enumerates resolution as 480P, 720P and 1080P, with 1080P as the default — 4K is not an accepted value. That settles what the API offers today; it is not a statement about what the model might support in a later revision.
How much does Wan 3.0 cost?
⚠️ No rate has been published. Neither the release log entry nor the API reference states pricing, and any per-second or per-clip figure quoted elsewhere did not come from Alibaba. We will record a rate here on the day one appears in an official source.
Will ImagineToVideo add Wan 3.0?
We have run Wan 3.0 in internal integration testing on pre-release access, but there is no launch to announce. Putting any model in the generator depends on a generally available API with published pricing and terms, and Wan 3.0 currently has neither. We track its status weekly on this page; the update log above records every change.
Wan 3.0 is gated. These models are not.
If you came here because you need 30-second clips with sound and multi-reference control, there are models you can run today. We list what each one actually does, with output we generated ourselves.