Wan 3.0 Source NotesField notes on the Wan 3.0 video API — sourced, dated, and corrected in public.

The model that is in preview and generally available at the same time

A small thing that turns into a real problem the moment you have to write a sentence with a date in it.

Wan 3.0's timeline, from Alibaba's own material:

Date What happened
2026-08-06 Public preview opens
2026-08-13 Availability widened on Model Studio
2026-08-24 General availability announced

And as of 2026-08-27, Alibaba's official English API reference still describes the model as "Currently in preview."

Both pages are live. Neither is a mistake in the sense of containing a wrong fact — the announcement is right that it shipped, and the reference page is stale rather than false. But if you are the person writing "Wan 3.0 is generally available" on a page that a reader can check, you are one click away from a source that says otherwise.

Why this is not pedantry

Documentation lag is normal and mostly harmless. It becomes harmful in one specific situation, which is exactly the situation anyone writing about a new model is in: a reader who checks your claim lands on the page that disagrees with you, and updates their trust in everything else you wrote.

The asymmetry is unpleasant. If you say "in preview" and it is GA, you look out of date. If you say "GA" and the vendor's reference says preview, you look like you made it up. Neither is recoverable by being right later.

The formulation that survives

Not "Wan 3.0 is generally available." Instead:

Alibaba announced general availability on 2026-08-24; the English API reference still carried the "Currently in preview" label when this was checked on 2026-08-27.

Longer, and it does three things the short version cannot:

  1. It is true on both pages. A reader who checks either one finds what you said.
  2. It survives the fix. When the reference page is updated, the sentence becomes history rather than becoming wrong.
  3. It tells the reader something useful — that this model is new enough for the vendor's own surfaces to be out of sync, which is itself a fact worth having if you are about to build on it.

The general rule I have ended up with: when two primary sources disagree, cite both and date them. Picking the one you prefer and presenting it as settled is the failure mode, and it is tempting precisely because it reads better.

Where else this model's sources disagree

Since the practice only means anything if you apply it consistently, here are the other places I know of in the same documentation set. All checked 2026-08-27.

enable_thinking. Four independent API gateways — fal.ai, Picsart, Modellix and WaveSpeed — document this parameter as required to process file and link inputs. Alibaba's Model Studio console emits it in its generated curl examples. And a full-text search of Alibaba's English API reference returns zero matches for thinking.

Four downstream sources plus the vendor's own tooling on one side; vendor silence on the other. The defensible sentence names both sides. The action, separately, is to set it to true anyway — because if it is required and you omit it, the request succeeds, bills, and quietly ignores your document.

Reference-video billing. The formula input video seconds + output seconds, rounded up, appears identically in evolink, APIXO and aivideoapi's docs. It is not stated in Alibaba's reference. But the official response object contains input_video_duration and output_video_duration as separate fields, and there is no other reason for that split to exist. Three secondary sources plus structural evidence from the primary one. Strong, still not a vendor statement.

wan3.0-video-prime. Costs about 53% more. QwenCloud's model list annotates it "(high-speed)". Its published output specifications — resolutions, durations, frame rate — are identical to the standard model. There is no vendor statement that it looks better, and no published comparison either way. "Premium quality tier" is an inference from the price, and inference should be labelled as such.

wan2.7-r2v's duration cap. The 2–10 second limit when a reference video is attached comes from apiyi and Together.ai — two sources that agree with each other and neither of which is Alibaba. The other rows in that table are from Alibaba's own model page. Same table, two evidence grades, and the table should say so.

The cheap version of all this

You do not need a methodology. You need one habit: write the date next to the number.

A specification is a fact about a moment. Six extra words convert a sentence that will silently rot into a sentence that stays true and merely becomes historical. Everything else above is downstream of that one habit.

Release timeline from Alibaba Cloud's announcement material and TechNode's 2026-08-24 report; "Currently in preview" from Alibaba Cloud's English wan3.0-video API reference, read 2026-08-27.

I keep a spec table with each row linked to its source and dated, including the rows where the sources disagree and the rows that are unflattering. Disclosure: wan-3.run is my site; it runs this model in a browser tab, and the first clip on it is free.