Wan 3.0 specifications

Wan 3.0 is Alibaba's all-in-one video generation model, published as wan3.0-video and wan3.0-video-prime. It generates any whole number of seconds from 2 to 30 at 30 fps, in 480P, 720P or 1080P, with speech, effects and music produced in the same pass as the picture.

This page exists because a large share of the material ranking for this model gets those numbers wrong. Every figure below is from Alibaba's own documentation — the Wan 3.0 API reference and the create-task schema — and every one of them can be checked in a click. Read 2026-08-26.

We run a third-party interface to this model, so read the page knowing that. The reason to correct the record rather than repeat it is that a specification is the one thing on a page like this a reader can verify before spending anything.

Five things Wan 3.0 does not have

These are the recurring inventions. Each of them appears in pages currently ranking for this model, and none of them is in the parameter table.

The identity lock people describe is not there either: nothing carries a subject across sessions. Reference material is per-request.

Output

With video input, input seconds and output seconds share the same thirty-second ceiling. A fifteen-second source clip leaves fifteen seconds of output, and asking for more is rejected before generation. duration: -1 does not escape this.

Input

Prompts run to 20,000 characters, in Chinese or English. Past that the prompt is truncated silently rather than rejected.

The two families are mutually exclusive

reference_image, reference_video, reference_audio, file and link are the reference family. first_frame and last_frame are the keyframe family. A request carrying both is rejected with InvalidParameter:

The two modes are mutually exclusive. Do not pass reference_xx and
first_frame/last_frame at the same time.

This is the single most common way a well-formed request fails, and it fails after queueing rather than at submission.

Citing an attachment in the prompt

In reference mode the prompt can point at an attachment by number — Image 1, Video 1, Audio 1 — numbered by position within its own type, so images and videos each count from one. A citation that does not match the documented form binds to nothing; the request still runs and the model invents the subject instead.

Accepted formats

Attachments are passed as a public HTTP(S) URL, an OSS URL, or Base64.

Task lifecycle

Generation is asynchronous. Submitting returns a task_id; the result is polled.

PENDING → RUNNING → SUCCEEDED | FAILED

The status enum has six values: PENDING, RUNNING, SUCCEEDED, FAILED, CANCELED, UNKNOWN.

UNKNOWN means expired, not failed. A task_id is valid for 24 hours; after that a query returns UNKNOWN. Treating it as a pending state is how an interface ends up spinning forever on a job that finished yesterday.

video_url also expires 24 hours after the job. Anything you intend to keep has to be copied to your own storage inside that window. This site does that automatically; if you are calling the API yourself, it is on you.

There is no cancel endpoint. CANCELED exists in the enum and no documented call moves a task into it. A thirty-second 1080P job, once submitted, runs.

Concurrency is the real constraint. Two slots, and a long job holds one for minutes.

Pricing, as Alibaba publishes it

Alibaba's list price is per output second, and the three tiers are a straight 1×/2×/4× ladder: $0.05 at 480P, $0.10 at 720P, $0.20 at 1080P. Converted to the unit leaderboards use, that is $3.00, $6.00 and $12.00 per minute.

Two things people get wrong about the bill:

• Reference video is billed. Its seconds are charged at your output rate, on top of the output seconds. Reference images, reference audio, documents and links add nothing.
• Audio is free either way. Picture and sound are generated in one pass, so turning sound off saves nothing.

What this site charges is a separate question, on the pricing page, and it is not Alibaba's price list.

Where Wan 3.0 sits against other models

On the Artificial Analysis text-to-video arena, With Audio board, snapshot 2026-08-25, Wan 3.0 leads on Elo at 1,240.

The caveat belongs in the same breath: that placement rests on 5,728 comparisons, the smallest sample in the top five, its margin over second place is three Elo points against a confidence interval near nine, and the board itself states its rank as a range of 1–2 rather than a rank of 1. "Joint top" is supportable. "Clearly the best" is not.

Head-to-head parameter tables against the models Wan 3.0 actually loses briefs to are in the comparisons: Veo 3.1, Kling 3.0, Seedance 2.5, MiniMax H3 and the previous generation, Wan 2.7.

Model identity

Wan 2.7 is the previous generation and a different set of four endpoints. Wan 2.2 is the last version in the family with published weights. Tongyi Wanxiang is Alibaba's own name for the family, and pages that treat it as a separate product are describing the same thing twice.

Questions

Does Wan 3.0 output 4K?

No. Three resolutions are published — 480P, 720P and 1080P — and 1080P is the ceiling. Pages listing a 4K tier are repeating a claim that is not in the parameter table.

How long can a Wan 3.0 clip be?

Any whole number of seconds from 2 to 30, in one pass rather than stitched. With video input, the input seconds count against the same 30-second ceiling.

Is Wan 3.0 open source?

No. There are no published weights, no checkpoint and no ComfyUI node. The most recent open-weight release in the family is Wan 2.2, under Apache 2.0.

What frame rate does Wan 3.0 output?

30 fps. The response reports it as usage.fps. Comparison tables listing 24 fps are describing a different model.

Does turning audio off make a clip cheaper?

No. Picture and sound are generated in the same pass and the documentation states that the audio switch does not affect pricing.

Can I send reference images and a first frame in the same request?

No. The reference family and the keyframe family are mutually exclusive, and a request carrying both is rejected with InvalidParameter after it has queued.

Why does a finished job come back as UNKNOWN?

Because the task_id expired. It is valid for 24 hours, and a query after that returns UNKNOWN — which means expired, not failed. The video_url expires on the same clock, so anything worth keeping has to be copied out inside the day.

Does Wan 3.0 have a director mode or a shot count?

No. There is no shot-count parameter. Multi-shot is a way of writing the prompt, so the number of cuts you ask for is a request rather than a contract.