Wan 3.0: What the Long Take Can Do, How to Set It Up, and What We Saw Testing It

At a glance

What Wan 3.0 changes

The headline change is length. Earlier Wan releases were built around shorter clips, and Wan 3.0 raises the ceiling to 30 seconds in one generation. The API reference describes it as a reference-based model: images, videos and audio can be passed in together, and the output can carry sound.

For a creator, that changes what a first attempt can be. A scene that used to need several short generations and a rough join can, in principle, be tried as one take. Whether a given scene holds together for the full length is something each person has to check on their own material. The published figure is a ceiling, not a promise about every prompt.

The Wan 3.0 feature set

One long take, up to 30 seconds

The API accepts durations from 2 to 30 seconds, at 30 fps. The default is 5 seconds. Setting the duration to -1 turns on smart duration, which lets the model choose a length from the prompt. We have not run a 30-second generation, so we cannot describe how a longer take holds together. What our 5-second clip does show is covered in the test footage section below.

Sound in the same generation

The audio parameter defaults to true, which means a generated video includes an audio track. The file we tested has a stereo AAC track. We did not run a listening review, so this article does not judge the quality of the sound. The web version we publish has its audio removed, because the page plays clips muted and the file is smaller without it.

References for images, video and audio

Reference-based generation accepts up to 10 images, 5 videos and 5 audio clips. Reference videos must be 1 to 15 seconds each, and the combined length of all reference videos cannot exceed 15 seconds. Reference audio follows the same pattern. Audio is passed as a reference in reference mode and can be combined with reference images and reference videos. The first-frame and first-last-frame modes take images only.

The product card says these references let Wan 3.0 follow one character through a whole shot. The API documentation explains how inputs are passed. It does not say how well a character holds across 30 seconds, and we have not tested that claim.

What our test clip shows

We generated the clip in VibePaper on 14 October 2026 as a 16:9, 1080p, 5-second video, from a text prompt that asked for one unbroken take through a lantern-lit night market in the rain, ending on a street musician. The first frame is a wide view of food stalls under red lanterns, with a chef working over open flames and steam rising from a pot. A mid-clip frame moves in among the crowd, with people under umbrellas and stall signs in the foreground. The last frame sits in a stone archway, with a cellist seated at the centre and lanterns overhead.

Real footage from our test

The clips below are from the same generation. The web version is 720p and muted, re-encoded from the original 1080p file. The stills are frames taken from the same clip, at the same size as the video.

HOVER TO PREVIEW

How to make a Wan 3.0 clip in VibePaper

Where each control lives

The settings row sits under the prompt box, beside the Generate button. The captures below follow it from left to right: the model, the aspect ratio, the resolution and the duration. They are taken from the English Paper Gen interface.

Limits and open questions

A feature list tells you what a model is meant to do. The limits decide whether it fits your work, so here is what we could and could not confirm.

• We have not generated a 30-second clip. The published ceiling is 30 seconds, but our only test result is 5 seconds, so we cannot say what a longer take looks like.
• Reference length is capped in total. Reference video and reference audio each have a 15-second combined limit per request, so a long reference set has to be trimmed first.
• Some modes take less. First-frame and first-last-frame modes take images only, and they do not accept audio.
• The Generate button did not show a price for Wan 3.0 in our check, and the API reference we read does not list one, so this article does not quote a price. Check the price in the product before you submit.
• The sound quality and the lip or voice behavior of Wan 3.0 are not judged here. We ran one test, and we did not listen to it in a controlled review.

A prompt checklist for your first attempt

One unbroken shot, no cuts. The camera moves at walking height through a narrow night market in the rain, past a grill with open flames and rising steam, and stops in a stone archway where a cellist plays. Warm lantern light against blue rain, wet paving that reflects the lanterns, photorealistic, no text.

Questions we get about Wan 3.0