Wan 3.0 API

Wan 3.0 is the new flagship video model from Alibaba's Tongyi Lab, and its pitch fits in one sentence: thirty seconds of 1080p video with sound, from almost anything you can hand it. A text prompt, a first frame, a first and last frame, up to ten reference images, short reference clips and audio, an existing video to edit or extend, and, for the first time in the Wan line, a document or a web page. All of it goes through one model rather than a shelf of task-specific ones.

The thirty seconds matter. Many video models of this period produce between 5 and 15 seconds per generation, so anything longer has to be stitched from several clips, and the joins show: a face shifts slightly, a jacket changes colour, the light jumps. Wan 3.0 produces a full 30-second take in one pass, with dialogue, music and sound effects generated alongside the picture.

Alibaba opened a public beta on 6 August 2026 and launched the model officially on 24 August. It comes in two tiers that share the same capabilities: the standard Wan 3.0 (wan/3-0-video on Unicorn API) and Wan 3.0 Video Prime (wan-3-0-video-prime), a faster version with the same feature set.

Where Wan 3.0 comes from

Wan is the video line of Tongyi Lab, the Alibaba group that also builds the Qwen language models. The family made its name as an open project: Wan 2.1 and then Wan 2.2, released on 28 July 2025, both shipped their weights under the Apache 2.0 licence and became a staple for people running video models on their own GPUs.

From Wan 2.5 onwards the flagship line went closed. Wan 2.5 added native audio in September 2025, Wan 2.6 pushed clips to 15 seconds in December, and Wan 2.7, in April 2026, split the work across separate models for text-to-video, image-to-video, reference-to-video and editing. Wan 3.0 continues the closed approach. There are no weights on Hugging Face, GitHub or ModelScope and no open licence; access runs through Alibaba Cloud Model Studio, Qwen Cloud, Alibaba's own apps and API providers such as Unicorn. Anyone who needs a Wan model they can download still has to go back to Wan 2.2.

Alibaba announced the launch the day after closing a large share sale whose proceeds are earmarked for AI infrastructure, and says Wan 3.0 is already in use for short dramas and film, advertising, tourism promotion and music videos.

What changed from Wan 2.7

One model instead of four. Wan 2.7 was a family: wan2.7-t2v, wan2.7-i2v, wan2.7-r2v and wan2.7-videoedit, each with its own limits. Wan 3.0 folds text-to-video, first-frame and first-and-last-frame animation, all-in-one references, editing and extension into a single model with a single set of rules. In practice that means you can mix inputs that used to need different endpoints, for example a character from a reference image, a voice from a reference audio clip and a scene from a text prompt.

Twice the length. Wan 2.7 text-to-video and image-to-video topped out at 15 seconds, and its reference and editing models at 10. Wan 3.0 accepts any length from 2 to 30 seconds in every mode. When you pass a video as input, the input and output together must stay within 30 seconds.

Documents and web pages as input. This is new for Wan and unusual anywhere. Wan 3.0 accepts one file (Word, Excel, PowerPoint, PDF, plain text, Markdown or Apple iWork formats, up to 100 MB and 50 pages) or one public web page, reads it and turns it into a video. The obvious uses are product explainers from a spec sheet, a summary clip from a slide deck, or a promo built from a landing page.

A draft tier. Wan 2.7 rendered at 720p and 1080p only. Wan 3.0 adds 480p, which is good enough to judge motion, timing and framing before committing to the final render.

More natural people. Alibaba says Wan 3.0 produces "diverse, lifelike human faces (no more same-face AI people)" with "natural micro-expressions", and claims more precise consistency for characters, props, spaces and style across reference-driven clips. Coverage of the launch also highlights better instruction following, shot consistency and audio quality.

Smarter defaults. An adaptive aspect ratio picks a frame shape from the inputs and the prompt, and a smart-duration setting lets the model choose the clip length itself.

Wan 3.0 or Wan 3.0 Video Prime

Alibaba's description of Prime is short: the "accelerated version" of Wan 3.0, with "capabilities aligned to the standard version" and "significantly faster generation speed while maintaining high-quality output". Both tiers take the same inputs, run to the same 30 seconds and offer the same resolutions. Alibaba has not published a speed figure, so the gain should be measured on your own workload.

What Prime offers, then, is time. Pick standard Wan 3.0 for almost everything: batch jobs, overnight renders, drafts and any workflow where a person reviews the result anyway. Pick Prime when waiting is the problem, such as reactive social content, a live event, or an app where a user is staring at a progress bar. On Unicorn, Prime defaults to 1080p while the standard tier defaults to 720p, so check the resolution before sending a batch. Prime also accepts -1 as a duration, which hands the choice of length to the model.

How Wan 3.0 works, as far as Alibaba says

Alibaba has not published a technical report, parameter count or architecture for Wan 3.0, and nothing about its training data. What it documents is how the model is driven. Reference inputs are numbered by type and order, so the prompt can say "Image1", "Video1" or "Audio1" and point at a specific asset. An edit is a reference video plus an instruction (Alibaba's example: "Transform the entire scene into a clay art style"); an extension is a reference video plus a prompt that says to continue it forwards or backwards. Prompts may be in Chinese or English and run to 20,000 characters, long enough for a full shot list or script. Output is 30 fps MP4.

Wan 3.0 specs at a glance

What Wan 3.0 is good at

Long single takes. A 30-second clip is a complete social post, a full product walkthrough or a short scene with a beginning, middle and end. Because it is one generation, the face, wardrobe and lighting hold together in a way stitched clips rarely do. One early hands-on review called the face consistency "genuinely good, much better than the melting-face experience" of older models.

Text-to-video with sound. In Artificial Analysis's blind-vote arenas, Wan 3.0 went straight to the top of the video-editing leaderboard with audio, at 1,189 Elo from more than 5,000 votes in August, and sits at or near the top of the text-to-video leaderboard with audio. That places it level with or ahead of rivals such as Seedance 2.5 and MiniMax H3, within the margins of error.

Reference-heavy scenes. Testers who fed it character sheets, locations and props found it combined them into one clip "surprisingly" well, which makes it a practical tool for series work with recurring characters and sets.

Screens and products. One hands-on review reports clean results on interface walkthroughs, with legible on-screen text and smooth camera moves, which together with document input makes Wan 3.0 a natural fit for software demos and explainer videos.

Editing instead of regenerating. Restyling a clip, changing what happens in it or extending it from an existing video keeps the parts you liked and saves a full re-roll.

Weaknesses and limits

Lip sync against supplied audio. The most common complaint from testers is mouth movement that does not quite match a provided song or voice track, and one review saw a small lip-sync drift in the last seconds of a long clip.

Unrequested cuts. Wan 3.0 sometimes cuts to a new camera angle even when the prompt asks for one continuous shot. Reviewers saw it often enough to treat it as a habit of the model, not a fluke.

Physics and busy scenes. Motion looks smooth but can break physical logic. With several characters and many props, consistency slips: backgrounds change between shots and composited characters can look pasted in.

Text and audio texture. Alibaba itself says "audio texture and on-screen text rendering accuracy are improving but not yet where we want them." Expect the odd typo in on-screen text and the occasional invented detail, such as a product in the wrong colour. Voices are clear, but not yet a substitute for a professional voice-over.

No 4K, no weights. Output stops at 1080p, and the model cannot be downloaded, fine-tuned or run privately. Alibaba has also published no benchmark scores of its own.

Prompting tips and example prompts

Write like a director. Give the subject, the action in order, the camera and what we hear. With 30 seconds to fill, a short sequence of beats works better than one long adjective-laden sentence, and putting spoken lines in quotation marks makes it clear what is dialogue.

For a talking scene from text alone:

An elderly fisherman mends a net on a wooden pier at dawn. His granddaughter runs up with a thermos; he looks up, smiles and says, "The tide waits for no one." Handheld medium shot. Gulls, lapping water, creaking boards.

For a longer single take, spell out the beats and say you want no cuts:

One continuous shot, no cuts. The camera follows a cyclist through a narrow market street: past fruit stalls, under strings of lanterns, out onto a bridge over a river, where she stops and looks at the sunset. Street chatter fading into wind and water.

For reference mode, name each asset by its number:

Image1 sits at the counter of the café from Image2 and orders a coffee in the voice of Audio1. The barista laughs and slides the cup across. Warm afternoon light, slow push-in.

For an edit, pass the clip as a reference video and describe only the change:

Transform Video1 into a hand-made clay animation style, keeping the camera movement and timing.

Two habits pay off. Draft at 480p, then re-run the winning prompt and seed at 720p or 1080p. And if the model cuts when you want one take, say "one continuous shot" explicitly, and check long clips for stray cuts before publishing.

How Wan 3.0 compares with other video models on Unicorn

Against Wan 2.7. Wan 2.7 Text to Video stops at 15 seconds and has no 480p tier, document input or unified reference mode. Wan 2.7 Image to Video is in the same position. For new work, Wan 3.0 is the better default.

Against Seedance 2.5. Seedance 2.5 also runs up to 30 seconds, and one detailed side-by-side review judged it the better choice for character-driven, cinematic work where lip sync and continuity matter most. Wan 3.0 is the workhorse for volume, supporting footage, editing and document-driven explainers.

Against MiniMax H3. MiniMax H3 Text to Video renders at 768P and sits close to Wan 3.0 in the text-to-video arena. Pick Wan 3.0 when you need 30-second takes, document input, or editing and extension in the same model.

Against Kling 3.0 Omni and HappyHorse 1.1. Kling 3.0 Omni offers clips up to 15 seconds and a 4K option, useful when resolution matters more than length. Alibaba's own HappyHorse 1.1 Text To Video is the other in-house option to test against Wan 3.0 on the same brief.

Getting started

Open the Wan 3.0 playground, write a prompt or upload a first frame, choose length, resolution and aspect ratio, and generate. Prime is in the same studio as Wan 3.0 Video Prime.

For code, create a key on the API keys page and submit a job to /v1/jobs with the model id and a prompt, plus duration, resolution and any image, reference or document inputs. The job returns a video URL when it finishes; the API docs cover authentication, polling and webhooks.

Verdict

Wan 3.0 is the most complete video model Alibaba has shipped: one model for generation, references, editing and extension, 30-second takes with sound, and the novelty of turning a document or web page straight into video. It still slips on lip sync, physics and on-screen text, and it is closed. But with arena results that match its strongest rivals and a 480p tier for drafting, it is an easy default for long-form and high-volume video, with Prime held back for the jobs that cannot wait.