A large number of pages ranking for Wan 3.0 list a 4K output tier. The model does not have one. Its resolutions are 480P, 720P and 1080P, and 1080P is the ceiling.
The interesting question is not that the claim is wrong. It is where it came from, because the origin explains why it spread and why it will keep spreading.
Checked against Alibaba Cloud Model Studio's Wan 3.0 parameter table, 27 August 2026.
The nearest plausible source is a real specification, belonging to a real Alibaba model,
that is genuinely capable of 4096×4096 output — wan2.7-image-pro.
An image model. In the same family, with an adjacent version number, and a maximum dimension that rounds to "4K" in ordinary usage.
Once you see it the propagation path is obvious:
Nobody in that chain lied. The number is real. It is attached to the wrong product, and the wrongness is invisible because it is exactly the kind of claim a new flagship would make.
Once you know to look for cross-product contamination, the other durable errors about this model sort into a short list, and most of them come from one of two neighbours: the image models in the same family, or a competing video model with similar positioning.
| Claim in circulation | Actual | Where it plausibly comes from |
|---|---|---|
| Up to 4K output | 480P / 720P / 1080P | wan2.7-image-pro, an image model, at 4096×4096 |
| 24 fps | 30 fps | The general assumption that video models are cinematic-rate |
| 21:9 supported | adaptive, 16:9, 4:3, 1:1, 3:4, 9:16 | A competing model that does offer a cinema ratio |
| Open weights available | Closed; last open flagship is Wan 2.2 | True of the family's earlier generations |
| 4–15 second clips | 2–30 seconds | A different vendor's video model |
| 7,000-character prompt limit | 20,000 characters | A different vendor's video model |
| 9 reference images / 3 clips / 3 audio | 10 images / 5 clips / 5 audio | A different vendor's video model |
Read down the right-hand column and the pattern is that every wrong number is a correct number about something else. That is what makes them durable. A fabricated spec gets challenged; a misfiled one gets corroborated, because the person checking finds the figure somewhere real.
One question, applied to every figure before you publish it:
Which exact model identifier is this number attached to on the page I read it from?
Not the family. Not the brand. The identifier — wan3.0-video, wan2.7-image-pro,
wan2.7-t2v. If the page you are reading does not name one, you cannot answer the
question, and the figure should not travel.
This catches all seven rows above. It costs one lookup per number. And it is the same discipline that makes a spec table maintainable rather than merely correct on the day it was written: if every row carries the identifier and the source, re-checking is mechanical.
There is a real cost to writing "no 4K". Pages that promise 4K rank for people searching for 4K, and a page that says the capability does not exist will lose some of those clicks.
The counter-argument is that the correction is the more defensible position over time. Someone who arrives believing there is a 4K tier and leaves knowing there is not has been given something they can act on, and they are one of the few readers in the category who was actually told a checkable fact. There is no version of this where the model gains the capability retroactively and the accurate pages look foolish.
It also compounds in a way the optimistic version does not. A page that states the unflattering specification, sources it, and dates it can be cited by anyone — including AI search systems answering "does Wan 3.0 support 4K", which need a source that actually answers no.
Every specification row on wan-3.run links to Alibaba's own parameter page and carries the date it was last verified, including the ones — no 4K, no 21:9, 30 fps not 24, no published weights — that cost traffic to state.
The faster way to settle a resolution claim is to stop reading about it. The tier selector in the text-to-video generator offers three values and no fourth, which is the same evidence the parameter table gives you and takes about four seconds to check.