How to use Wan 3.0: four routes in, and what each one asks of you

The question behind "how to use Wan 3.0" is usually not how do I write a prompt. It is where do I actually go, and that is genuinely confusing right now, because the model shipped as an API before it shipped as a product.

There is no single Wan 3.0 app. There are four routes in, they have very different entry requirements, and the difference between them is mostly paperwork rather than output. The model is the same at the end of all four.

Every parameter named below is from the Wan 3.0 API reference and the create-task schema, read 2026-08-25.

Route 1 — Alibaba Cloud Model Studio

The first-party developer route, and the one every other route is built on. You create an Alibaba Cloud account, pick a region, generate an API key and call the video synthesis endpoint.

Two things to know before you start, because both cost an afternoon if you learn them late. The model, the endpoint and the API key must all be in the same region — cross-region calls simply fail, so choosing a region is effectively choosing a key. And the call is asynchronous by design: you must send the async header, you get a task_id back rather than a video, and you poll for the result. Roughly every fifteen seconds is the documented rhythm, and a job usually lands in one to five minutes.

The domestic Chinese region has a further requirement worth naming plainly: it expects a verified Alibaba Cloud account, which for people outside China is often where this route stops. The international endpoints exist for exactly this reason.

Route 2 — Qwen Cloud

The same models under an international-facing wrapper, with wan3.0-video and its faster sibling wan3.0-video-prime both listed. Same async pattern, same parameters, same 2–30 seconds at 480P, 720P and 1080P, 30 fps, MP4.

Availability has been rolling out gradually rather than opening all at once, so whether you can sign up today depends on when you read this. Worth checking first rather than planning around.

Route 3 — Alibaba's own consumer surfaces

Alibaba does run first-party creative products, and Wan 3.0 has been arriving on them progressively rather than all at once — described in Alibaba's own material as rolling out to members. There is also an international web presence and a mobile app.

Two honest caveats, since a lot of English coverage gets this wrong in both directions. It is not true that there is no first-party consumer product; there is. It is also not true that you can simply sign up and start — several of these surfaces are built for the Chinese market first, with a Chinese interface and, in places, a Chinese phone number at registration, and access has been gated rather than open.

If you have an Alibaba account and patience, this is the most authoritative route. If you have a deadline, it is not the one.

Route 4 — a hosted front end

Someone else has done routes 1 or 2 and put a browser interface on top. You write a prompt and press a button. This is what most people searching "how to use Wan 3.0" actually want, and it is what this site is.

The thing worth checking is that a hosted service is genuinely running Wan 3.0, because the search results for this are noisy. Three questions answer it:

1. Does the result show a model ID? wan3.0-video or wan3.0-video-prime. Anything vaguer is a choice, not an oversight.
2. Do the capabilities match the real ones? A service claiming a resolution tier above 1080P, or a six-shot director mode, is describing a model that does not exist.
3. Do the limits match? 2 to 30 seconds, 480P / 720P / 1080P, ten reference images, five reference videos, five reference audio clips, one document or one link.

That third one is the most reliable tell, because the limits are oddly specific and hard to fake convincingly.

Your first clip, whichever route you took

The settings are identical everywhere, and four of them decide almost everything.

Resolution — set it. If you leave it unset, Wan 3.0 generates at 1080P, which is four times the price of 480P. Draft at 480P; it is the same model with fewer pixels, and framing, motion and sound all read correctly there.

Duration — name a number. 2 to 30 seconds, any integer. You can pass -1 and let the model choose, but then the cost is unknown at submit time. Start at five seconds while you are finding the shot.

Audio — on unless you have a reason. Sound is generated in the same pass as the picture, and toggling it does not change the price. There is no money to save by muting.

Aspect ratio — or let it adapt. 16:9, 4:3, 1:1, 3:4, 9:16, or adaptive. There is no 21:9; if a comparison table told you otherwise, it was describing a different model.

Then the prompt. One thing beats every trick: describe how the shot changes across its length, not how it looks in one frame. "A courier in a wet yellow jacket shoulders a door open and steps into a neon alley, camera pushing in slowly, rain on the soundtrack" gives the model a subject, a motion, a camera and an audio bed. "Cinematic urban scene" gives it nothing and it will invent all four.

Two settings that surprise people:

• Prompt expansion is on by default. It genuinely helps a one-line prompt and it will rewrite a long one you have carefully tuned. Turn it off once your prompt is exact.
• The model decides whether to cut. Left unspecified it will often cut between setups. Write "one continuous take" if you want an unbroken shot. More on controlling that here.

The five limits that shape a first session

The two 24-hour clocks are the ones that bite. A generated clip nobody downloaded is gone, and it does not appear anywhere as a failure — it was a success nobody collected. The other failure modes are here.

A first session that does not waste money

1. 480P, five seconds, audio on, one subject, one action, one camera move.
2. Read what came back. Was the framing right? The motion? The sound?
3. Change one thing and run it again. Still 480P.
4. When the five seconds are right, extend the prompt into beats and run the full length — still at 480P.
5. Run the winner once at 720P or 1080P with the identical prompt.
6. Download it before the link expires.

Four 480P drafts and one 1080P final costs about the same as two 1080P attempts, and you will have learned four times as much. The most expensive habit in AI video is not picking the wrong model; it is deciding what the shot is at full resolution.

Questions

Is there an official Wan 3.0 app?

Alibaba has first-party consumer surfaces and access to Wan 3.0 on them has been rolling out to members rather than opening at once, with several built for the Chinese market first. The developer routes — Model Studio and Qwen Cloud — are the reliably available first-party options.

Do I need an Alibaba Cloud account?

For routes 1 and 2, yes, and for the domestic Chinese region a verified one. A hosted front end exists precisely so that you do not.

Is Wan 3.0 free to use?

The model is metered from the first second — Alibaba publishes $0.05, $0.10 and $0.20 per second at 480P, 720P and 1080P. Some hosted services cover a first clip themselves; nobody is getting the model for free, and a site claiming unmetered access to it is describing something else.

How long does a generation take?

Usually one to five minutes, longer at thirty seconds. Two jobs run at once with up to fifty queued, so a batch waits rather than fails. There is no cancel call in the published API, so a long job submitted is a long job you are committed to.

Can I run Wan 3.0 on my own machine?

No. It is closed, with no published weights and no ComfyUI node. Alibaba's last open-weight video flagship is Wan 2.2, which is downloadable and a generation behind.

What should my very first prompt be?

Something small and specific: one subject, one action, one named camera move, five seconds, 480P. You are calibrating your expectations against the model's defaults, and that is cheaper to do on a simple shot than a complicated one.

---

Route 4 is this page. Write one prompt and the first clip is free — 480P, three seconds, with the Wan 3.0 model ID printed beside the result so you can see which of the four routes you actually got.