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

Two model IDs is the entire version surface of a closed video model

If you came to Wan 3.0 from the open-weights side of this field, the reproducibility story is going to feel thin, and it is worth being precise about exactly how thin rather than vaguely uneasy about it.

Here is the whole surface you can pin:

What you'd want to pin	What Wan 3.0 gives you
Weights	Nothing. Closed, nothing to download
Checkpoint / revision hash	Nothing
Seed	Not a published parameter
Model identity	Two IDs: wan3.0-video and wan3.0-video-prime
The artefact itself	24 hours, then the URL and the task ID both stop resolving

That is it. Two string literals and a one-day window.

What that rules out

Bit-exact reproduction is off the table. Not "hard" — unavailable. There is no parameter you can set that makes today's request return today's output tomorrow, and there is no revision identifier that would let you say which version of the model produced a given clip, beyond the two-way choice of ID. If the vendor updates the weights behind wan3.0-video, your requests change and nothing in the response tells you.

So does after-the-fact verification. Combine that with the 24-hour expiry and a clip you generated last month is unauditable in a strong sense: you cannot re-derive it, and the vendor no longer holds the record. Whatever provenance you need has to be captured by you, at generation time.

What to record instead, since you cannot pin

Since the platform will not do it, the burden moves to your side. The minimum useful record, written at submission:

• the exact prompt string, before any templating your own code applies and after — they diverge more often than people expect
• the model ID, as a literal, not as a variable name
• every generation parameter you set, including the ones you left at default (write the resolved value, not "default")
• the usage block from the response, which is where the returned duration and frame rate actually come back
• your own UTC timestamp

That record is not reproducibility. It is provenance: you cannot regenerate the clip, but you can say truthfully what produced it. For most commercial uses that is the requirement anyway — the question is "what made this" for an audit or a rights query, not "make it again".

If you actually need reproducibility, the answer is a different model

This is the honest conclusion, and it is not the one a site like this is supposed to reach: if bit-exact reproduction is a hard requirement, Wan 3.0 is the wrong tool and no amount of careful logging fixes it. The last flagship in this family with published weights is Wan 2.2, and open weights are the only thing that gives you a pinnable artefact — you hold the checkpoint, so the checkpoint cannot change under you.

The trade is real and it is large: Wan 2.2 tops out at 5 seconds and 720P with no audio, against 30 seconds and 1080P with speech, effects and music in the same pass. Whether that trade is worth it depends entirely on whether "I can prove what produced this, six months from now, by regenerating it" is a requirement or a preference. It is worth deciding that on purpose rather than discovering it during an audit.

Sources

The two model IDs, the closed-weights status, the 24-hour expiry, and the Wan 2.2 comparison figures are all documented parameters. They are collected with citations and check dates on the Wan 3.0 specification table — including the four claims in circulation that are not in any primary source.

Disclosure: this site is written by the people who run a third-party browser interface to this model. We are not affiliated with Alibaba or Alibaba Cloud.