Wan 3.0

What Wan 3.0 is good at — and what it's not

Wan 3.0 is Alibaba's answer to the question every video model has been dodging: what happens when the shot needs to last longer than a few seconds? Most flagships cap a single generation somewhere between five and fifteen seconds, which means anything resembling a scene gets assembled in an editor from clips that never quite match — the face drifts, the light shifts, the room changes shape at the cut. Wan 3.0 renders up to thirty seconds in one pass, and that single change reorganizes how you plan a shot. A continuous camera move through a space, a complete thirty-second ad read, a one-take product walkthrough: these stop being edit problems and become prompt problems.

The second thing that sets it apart is All-in-One Reference. Reference systems usually take one kind of input — some models take images, a few take a clip, a rare one takes a voice. Wan 3.0 takes all three in the same request: up to ten reference images, five reference clips and five reference audio tracks, capped at fifteen seconds of reference video and fifteen seconds of reference audio. In practice that means a whole brand kit or a whole character sheet goes in at once, and the identity holds for the length of the clip rather than for the first few seconds. On ZOOOP the routing is automatic — attach any reference and the request goes to the reference pipeline; leave them empty and the same model runs plain text-to-video.

Two more capabilities matter day to day. First and last frame control lets you pin both ends of a clip and have Wan generate the motion between them, which is how you land an exact ending that cuts cleanly into the next shot. And native synchronized audio — dialogue, ambience, music — is generated with the picture in the same pass and arrives as a stereo track on the file, with the audio toggle costing nothing either way.

Where it's weaker: it is not open-weight. Wan 2.7 shipped under Apache 2.0 with published weights, and Wan 3.0 did not — no weights, API only. For a pipeline that needs self-hosting or open provenance, that is a hard stop and 2.7 is still the answer. Its 1440p tier is an enhanced upscale, not a native render — a real 2560×1440 file, good for large-screen delivery, but a model with a native high-resolution pass will hold detail better. There is no negative prompt, so exclusions have to be handled by describing what you do want. And on image-to-video the output follows the aspect ratio of your input frame — there is no ratio control on that path, so crop the source before you generate.

A reasonable mental model: reach for Wan 3.0 when the shot is long, or when consistency against real reference material is the thing that decides whether the take is usable. For open weights and instruction-based edits, Wan 2.7. For top-end single-shot fidelity, Veo 3.1. For quick throwaway drafts on the same family, Wan 2.6 Flash. And when you are still deciding what the shot is, run the Prime tier and switch back for the final render.