Short notes on the operational side of generative video: what happens to a job after you submit it, what the platform keeps, and for how long.
These notes are written by the people who run wan-3.run, a third-party browser interface to Alibaba's Wan 3.0 video model. We are not affiliated with Alibaba or Alibaba Cloud, and we say so on every page here and on the product itself.
That is a commercial interest, and it is why the pages here are built the way they are: every factual claim about the model traces to Alibaba's own API reference, with the date it was last checked. You should not have to trust a vendor-adjacent site — you should be able to check it in a click, and where we could not verify something, it does not appear.
Because they are the ones missing. There is no shortage of material on writing prompts for this model. There is almost nothing on the parts that decide whether an integration holds up: result expiry, batch semantics, concurrency, and what can be pinned for reproducibility. Those are the questions that arrive in week three, after the demo works.
A lot of what currently ranks for this model is wrong in specific, checkable ways — a 4K tier that does not exist, a 21:9 ratio that is not in the list, 24 fps when the output is 30, and "open weights coming soon" with no vendor statement behind it. The Wan 3.0 specification table exists to be a thing you can point at instead: every published parameter, each one sourced and dated, including the unflattering ones.
If you find an error on any page here, it is worth reporting — a specification page that is wrong is worse than no page at all.