Happy Wan

Reference

Wan 3.0, the model that actually shipped

What Wan 3.0 is, what thirty-second clips with native audio change about prompting, and why it does not answer the Happy Wan question.

Wan 3.0 is the model that exists while Happy Wan is still a rumour. If you are waiting for the name, this is what you can actually shoot with today, and how its features change the way you write a prompt.

What Wan 3.0 is

Alibaba introduced Wan 3.0 on 24 August 2026. It is an all-in-one video model: text, images, audio and video go in as direction, and a clip with picture and sound comes out. The reported ceiling is 30 seconds per clip, and API resellers advertise 1080p output on the standard and Prime tiers. Earlier Wan releases were shorter and silent — Wan 2.1 in February 2025 was open-weight and text-to-video first, and audio arrived later in the 2.x line.

That single change — thirty seconds with sound — is what makes the prompting advice on this site different from the advice you will find for older models.

What changes when a clip is thirty seconds

A five-second clip is a gesture. A thirty-second clip is a scene, and scenes have rules that short clips can ignore.

  • One action is no longer enough. Thirty seconds needs a beginning, a middle and an end: something is set up, something happens, something settles. Write all three into the prompt as visible events, not as feelings.
  • Sound becomes part of the structure. When audio is generated with the picture, the moment a sound happens anchors the edit. Write “the base of the mug meets the wood” rather than “add foley later”, and the model has something to sync to.
  • Continuity failure gets more visible. Small hand and prop drift that nobody notices at four seconds is obvious at thirty. One subject, one prop, no second character, and no scene change inside the take.
  • You can cut inside the model. Two or three shots in one generation means you can plan a sequence; carry the last frame of one shot into the first frame of the next to keep faces and props stable.

A prompt shape that fits the model

[Shot] Medium shot, slow push in, one continuous take.
[Subject] A potter in an indigo apron holding a cobalt-blue mug by its single handle.
[Action] She lowers the mug onto a wooden table; the base meets the wood and she lets go.
[Light] Soft window light from camera left, falloff toward the back wall.
[Look] Cinematic, natural colour, 24 fps motion blur, focus held on the subject.
[Duration] 15 seconds, 16:9.
[Audio] One soft clink of ceramic on wood, quiet studio room tone.
[Avoid] No text overlays, no extra people, no camera shake, one light source only.

The home page builder writes this structure for you from seven controls, and the prompt library has ten finished examples.

What Wan 3.0 does not settle

Wan 3.0 does not answer the Happy Wan question. It does not tell you whether Alibaba’s next video model will be called Happy Wan, whether it will merge the reference-image control of HappyHorse 1.1, or when either will be available at what price. Those are open questions, and the status line on the home page is where we record the answer — with a date — once a source exists.

What Wan 3.0 does do is raise the floor. Character consistency, sound, and clip length are no longer reasons to wait for a new model name. If you can shoot the scene with what shipped in August, the name is a bookmark, not a blocker.

Quick answers

Is Wan 3.0 free? There is no free tier run by this site, and this site sells nothing. Alibaba’s own platform and the API resellers each set their own pricing; check theirs, and compare per second rather than per clip.

Can I use Wan 3.0 output commercially? That depends on the licence of the platform you generate on, not on this page. Read the vendor’s terms before you publish.

Does Wan 3.0 need a reference image? No, text alone works. Reference images are what keep a face or a prop recognisable, so use them for anything with more than one shot.

Where does Happy Wan fit in? Nowhere yet, officially. It is an unconfirmed name, and treating it as a released model is how most of the pages on this keyword get things wrong.