Wan 3.0 Video Generator: one model id, up to thirty seconds
Alibaba Cloud Model Studio publishes Wan 3.0 under the model id wan3.0-video, and calls it an All-in-One model because text-to-video, image-to-video, multi-modal reference and video editing all run on that one id without a model switch. This page runs three of those task types, quotes the official resolution, duration, reference, billing and rate-limit tables, and writes out every setting the model has that this workspace does not.
Write a prompt, or upload one frame, or upload one reference video, then set a duration, a resolution and a ratio, and read the point estimate before you submit. The official model also accepts up to ten reference images, five reference videos, five reference audio clips, one document or one public link, and a first-and-last-frame pair. Those inputs have no control here, and the difference is listed below rather than glossed over.
Alibaba Cloud Model Studio, built by the Tongyi Wanxiang team
Official model name
Wanxiang 3.0, written Wan3.0 throughout the documentation
Official model id, standard
wan3.0-video
Official model id, speed-optimised
wan3.0-video-prime
Lifecycle
Currently in preview, as stated at the top of the API reference
Announced
13 August 2026, in the official community blog post
Documentation last updated
Usage guide 11 September 2026, API reference 14 September 2026
Model type in the official table
All-in-One reference
Output resolutions
480P, 720P and 1080P, with 1080P as the documented default
Aspect ratio
16:9, 4:3, 1:1, 3:4 and 9:16, plus adaptive, which is the documented default
Duration without video input
2 to 30 integer seconds, default 5; -1 asks the model to choose a smart duration
Duration with video input
The input video and the output video together must not exceed 30 seconds
Frame rate and container
30 fps, MP4
Audio
Generated in the same pass and on by default; setting audio to false removes the track and does not change the bill
Multi-shot narrative
Native, in shots of 4 to 6 seconds marked with timestamps, up to the 30 second ceiling
Reference inputs
Up to 10 images at 20 MB each, 5 videos totalling 15 seconds, 5 audio clips totalling 15 seconds, and one document or one public link
Mutually exclusive inputs
Reference media, a document or a link cannot be combined with a first frame or a first-and-last-frame pair
Billing rule
Billed seconds are the input video seconds plus the output video seconds; a failed request is not billed
List price, Singapore region
0.05 US dollars per second at 480P, 0.10 at 720P and 0.20 at 1080P for wan3.0-video
Rate limit
5 requests per second and 5 concurrent tasks in every region the documentation lists
Result availability
The task id and the output URL both expire 24 hours after the task succeeds
Route id on this site
wan3.0-video, a platform identifier that happens to match the official model id
Documented by Alibaba Cloud, and offered on this page
The gap here runs one way and it is wide, so the two lists below are the point of this section rather than a footnote. Everything in the first column is a capability the official documentation registers for wan3.0-video. Everything in the second column is a control that exists in the workspace on this page today.
OFFICIAL
Registered for wan3.0-video in the official documentation
One model id covering text-to-video, first-frame, first-and-last-frame, multi-modal reference, document or link input, video editing and video extension, routed automatically by the type field and the prompt intent
Text-to-video from a prompt alone, with resolution, ratio and duration all free to set
Image-to-video from a first frame, with adaptive ratio recommended so the model matches the frame it was given
A first frame plus a last frame as the two ends of one controlled move
Up to 10 reference images at 20 MB each
Up to 5 reference videos totalling 15 seconds, at 100 MB per clip
Up to 5 reference audio clips totalling 15 seconds, at 15 MB per clip
A document of up to 50 pages, or one public web page, as the sole input
Native audio in the same pass, on by default and removable with audio=false without changing the cost
A smart duration, set by passing -1, which lets the model settle on a length instead of the caller
480P, 720P and 1080P output, with adaptive ratio and 30 fps MP4 as the container
Video editing and video extension, both driven by a reference video plus an intent word in the prompt
LIVE PLATFORM
Offered on this page today
Text-to-video from a prompt, image-to-video from one uploaded frame, and reference-to-video from one uploaded reference video
A 2 to 15 second duration selector with 5 seconds as the default, against an official range of 2 to 30
720p and 1080p in the resolution selector, against an official pool that also includes 480P
Five fixed aspect ratios with 16:9 as the default, and no adaptive option
One reference video per submission, with an accept list that includes WebM, which the official format table does not register
No first-and-last-frame route, no reference image upload, no audio input or audio switch, no document or link input, and no editing or extension mode
The distance is not entirely one way, and it is worth naming the two directions separately. The workspace offers one input the official format table does not register: the reference video accept list includes WebM, where the documentation lists mp4 and mov only. The workspace default resolution is also 720p where the documentation default is 1080P, which is a site-side choice rather than a model limitation. Everything else is a genuine absence. First-and-last-frame is closed on this route even though the model supports it, reference mode opens a video upload but exposes no image channel at all, and there is no control for audio, smart duration, documents, links, editing or extension. Read the official page as the ceiling of the model and the workspace as the floor you are actually working on, and do not read this page as evidence that the model lacks anything in the first list.
Model brief
What Wan 3.0 is, and why the model id matters
Wan 3.0 is the version of Alibaba Cloud Model Studio that the documentation spells Wanxiang 3.0 and writes as Wan3.0. The string that actually identifies it is different from both: the model id is wan3.0-video, with a dot inside the version number and a -video suffix, and the speed-optimised variant is wan3.0-video-prime. That dot is worth noticing, because this route is named wan-3-0 with hyphens, and the two strings do not contain each other in either direction. A page that presents the route name as the product identifier is describing this site's own naming rather than the model, and the API reference is unambiguous about which string a request has to carry. The API reference also states, at the top, that the model is currently in preview, and the official community blog post that introduced it is dated 13 August 2026, while the usage guide and the API reference were last updated on 11 and 14 September 2026.
The second thing to understand is what All-in-One means in the official wording, because third-party coverage routinely turns it into a list of separate models. The documentation says that the model covers all of its task types without switching model names, and that it routes automatically from the type field in the input media array together with the intent in the prompt. So text-to-video, first-frame image-to-video, first-and-last-frame image-to-video, multi-modal reference, document or link to video, video editing and video extension are not seven endpoints. They are one id, and the request decides which behaviour it gets. That has a practical consequence: a request that mixes reference media with a first frame or a last frame is rejected rather than merged, with the API returning an InvalidParameter error and the message that the two modes are mutually exclusive.
The third thing is the specification set, which is unusually explicit for this family. Output resolutions are 480P, 720P and 1080P, and the default is 1080P rather than the lowest tier. Aspect ratio can be set to 16:9, 4:3, 1:1, 3:4 or 9:16, or left on adaptive, which is the documented default and which matches whichever source material was supplied. Duration runs from 2 to 30 integer seconds with 5 as the default, and passing -1 hands the choice to the model as a smart duration. There is one constraint that third-party pages consistently miss: when a video is supplied as input, the input video and the resulting output video together must not exceed 30 seconds. Audio is generated in the same pass as the picture rather than added afterwards, it is on by default, and audio=false removes the track without changing what the request costs.
The fourth thing is the one most likely to be wrong in a summary written from the version number alone. Wan 3.0 does not simply inherit everything Wan 2.7 could do. The API reference records a specific regression: where the Wan 2.7 image-to-video series accepts driving audio alongside a first frame or a first-and-last-frame pair, the Wan 3.0 first-frame and first-and-last-frame modes do not, and the documentation tells callers who need audio-driven generation to use the all-modal reference mode with a reference image and a reference audio clip instead. That is an official statement that a newer generation is narrower in one respect, and it is the reason this page treats the official table as the source of truth for capability rather than the version number.
03 / Capabilities
What the official documentation establishes
Each point below is read from an Alibaba Cloud Model Studio page or from the official community blog post, not from a third-party summary of them.
01
One model id, seven task types, no model switch
The documentation is explicit that the model covers every listed task type without a caller switching model names, and that routing happens automatically from the type field in the media array plus the prompt intent. That removes the usual guesswork about which endpoint to call, and it also means a prompt that reads like an edit instruction is interpreted as an edit rather than as a generation, which is why the reference documentation separates the intent words rather than the identifiers.
02
1080P is the default, not the upgrade
The resolution parameter accepts 480P, 720P or 1080P, and the documented default is 1080P. Most video APIs default to the cheapest tier and treat the top tier as a paid upgrade, so a caller who does not read the parameter list will spend more than expected on every request. The ratio parameter behaves in the same direction, defaulting to adaptive so the output follows the proportions of whatever was supplied.
03
Thirty seconds, with a smart duration on top
Passing no duration gives 5 seconds. Passing an integer from 2 to 30 gives that many seconds. Passing -1 asks the model to decide, which the documentation calls a smart duration and which settles on 5 seconds as the fallback. The ceiling is not only a per-request limit: when a video is supplied as input, the input and the output share the same 30 second budget, so a long reference clip eats into how much can be generated.
04
Audio arrives with the picture, and switching it off is free
The model produces dialogue, sound effects and background music in the same pass as the video, without a separate dubbing step, and audio is on by default. Setting audio to false removes the track and leaves the price unchanged, because the billing rule counts video seconds rather than modalities. That makes the audio decision an editorial one rather than a budget one, which is not true of every route on this platform.
05
Multi-shot narrative is generated, not stitched
The documentation describes native multi-shot narrative up to the 30 second ceiling, with shots of 4 to 6 seconds and timestamps marking where the cuts fall. A sequence assembled this way is one generation rather than several clips joined afterwards, which matters for keeping the subject and the light consistent across a cut that nobody had to place by hand.
06
Reference media counts in three separate currencies
Reference images are counted in units of ten, at 20 MB each. Reference videos are counted both in units and in time, five clips totalling 15 seconds at 100 MB per clip. Reference audio is the same shape, five clips at 15 seconds and 15 MB per clip. The array order decides the numbering the prompt refers to, so the first reference video is Video 1 and the first reference image is Image 1, counted independently of each other.
07
A document or a public page can be the whole input
Passing a file or a link instead of reference media lets the model read a document of up to 50 pages, or a public page that needs no login, and turn it into video. The two are mutually exclusive with each other but compatible with reference media, and passing either one forces prompt extension on, because the request needs the model to write a usable prompt out of the source material before it starts generating.
08
Editing and extension run through reference video, not a separate model
Adding or removing elements, replacing a background, changing the light, rewriting a line of dialogue and extending a clip forwards, backwards or in both directions are all registered task types on this one id. Both require a reference video plus an intent word in the prompt, so neither is a one-click operation and neither works from a prompt alone. The extension is also bounded by the same 30 second budget as the input it extends.
09
The bill counts the seconds you supplied as well as the seconds produced
The documented formula is that billed duration equals the input video duration plus the output video duration. Supplying a ten second reference clip and generating a ten second result is therefore billed as twenty seconds, which is the opposite of the intuition that only the delivered file is charged for. Failed requests are not billed, so the risk attached to a long reference clip is a successful generation rather than an error.
04 / Scene ideas
Work that suits a two to thirty second generation
Each point below is read from an Alibaba Cloud Model Studio page or from the official community blog post, not from a third-party summary of them.
00:00:00SHOT 01Storyboard placeholder
Narration that has to carry its own sound
Because dialogue, effects and music are produced alongside the picture, a short scene can be finished in one pass instead of being sent to a separate audio step. That suits explainer inserts, character beats and any clip where the sound is part of the idea rather than a layer added afterwards.
00:00:02SHOT 02Storyboard placeholder
Sequences that need more than one shot inside one generation
The multi-shot mode produces cuts inside a single clip, marked with timestamps, which is how a thirty second piece can read as a scene rather than as a continuous take. It removes the step where several clips have to be matched for light and wardrobe because they were generated separately.
00:00:04SHOT 03Storyboard placeholder
Bringing a still product or portrait frame into motion
Uploading one frame and describing the movement keeps the composition the supplier already approved. The model matches the frame proportions when ratio is left adaptive, so a prepared image arrives on screen in the shape it was drawn in rather than being re-framed by the tool.
00:00:06SHOT 04Storyboard placeholder
Reusing an existing clip as the motion source
A reference video can drive the movement of a new generation, which is the practical route when the camera work already exists but the subject or the style has to change. On this page that route is the only video input, so it is also the closest thing here to an edit workflow.
00:00:08SHOT 05Storyboard placeholder
Turning a written document into a short piece
The official model accepts a document of up to 50 pages or a public web page as the entire input and writes its own prompt from the content. That fits training material, product notes and published articles, where the source text already exists and the work is turning it into something watchable.
00:00:10SHOT 06Storyboard placeholder
Editing an existing shot without reshooting it
Adding, removing or replacing an element, changing the light, or rewording a spoken line are registered capabilities on this model, driven by a reference video and an intent word. The value is in leaving the rest of the shot untouched, which is what a reshoot cannot promise.
00:00:12SHOT 07Storyboard placeholder
Extending a clip that is already right
Extension runs forwards, backwards or in both directions from a supplied reference video, and it respects the same input-plus-output budget. For material that has to run longer than one generation but must not look like two different pieces, it is the route the documentation registers.
00:00:14SHOT 08Storyboard placeholder
Testing a shot before committing to a long one
A two second generation costs a fraction of a thirty second one and answers the same questions about framing, motion and subject consistency. Running the short version first is the cheapest way to find out whether a description actually reads on screen.
05 / Prompt lab
Prompt starters for the routes on this page
Each example is written for a route that exists here, so it assumes either a prompt alone, one uploaded frame, or one uploaded reference video. Duration, resolution and ratio are set outside the prompt, and no example relies on reference images, reference audio, a document or a last frame, because none of those has a control on this page.
SCENE 01text to video
Single take, written from nothing
A text-to-video prompt that fixes the camera before it fixes the subject.
View prompt＋
One continuous shot, no cuts. A slow dolly to the right along a rain-wet street at dusk, camera held level at chest height and never raised. Neon signage reflects in the standing water and stays legible. One cyclist crosses from right to left in the middle distance and leaves frame. Keep the colour temperature warm against the blue of the sky. No text overlays, no cuts, no camera shake.
SCENE 02image to video
Product turn from a single frame
One uploaded frame, with the movement confined to what has to move.
View prompt＋
Hold the uploaded product exactly in shape, label position and the reflection underneath it. The camera does not move. The object rotates a quarter turn to the left at an even speed, then stops. Keep the label reading the same way throughout and keep the highlight along the top edge. Change nothing else in the frame and add no objects.
SCENE 03image to video
Portrait with one deliberate gesture
A small movement that is easier to judge than a large one.
View prompt＋
Keep the uploaded person unchanged in face, hair, clothing and the direction of the existing light. Add one slow turn of the head toward the camera and a single breath, then hold still. Do not move the camera. Keep the background layout and the shadow on the wall exactly as supplied. Nothing enters or leaves the frame.
SCENE 04image to video
Interior revealed by a forward push
Let one uploaded frame carry the whole description of a space.
View prompt＋
Preserve the uploaded interior in proportion, material and the position of every light source. Push the camera forward along the centre line at a constant speed, with no pan and no cut. Dust drifts through the shaft of light from the window on the left and a curtain settles at the far end. Keep the floor line and the ceiling height unchanged for the whole clip.
SCENE 05reference to video
Style transfer from an uploaded reference video
A reference video route, which is the one video input this page accepts.
View prompt＋
Treat the uploaded clip as the source of movement and timing, and keep its pacing exactly. Render the same action in a flat two-colour woodblock print style, with visible paper grain and hard edges. Keep the subject centred and the horizon level. Do not add transitions the source clip does not have, and do not change the direction of any movement.
SCENE 06reference to video
Camera move driven by a reference clip
Use the reference for motion only and describe the scene separately.
View prompt＋
Follow the camera path in the uploaded reference clip exactly, including its speed changes, and ignore its subject. Place a model car on a wet studio floor and let the reference camera orbit it from left to right at the same rate. Keep the reflections consistent with the orbit and keep the background as a seamless grey sweep. No cuts and no added titles.
COMPARE
Wan 3.0 against the Wan 2.7 rows in the same official table
Every value below is read from one table on one official page, the model overview that lists both generations side by side. Nothing here is a judgement about which model is better, and no third-party comparison is used as a source.
Wan 3.0 against the Wan 2.7 rows in the same official table
Item
wan3.0-video
wan2.7-t2v
wan2.7-i2v
Type label in the table
All-in-One reference
text-to-video
image-to-video
Output specifications
480P, 720P, 1080P
720P, 1080P
720P, 1080P
Duration range
2 to 30 seconds
2 to 15 seconds
2 to 15 seconds
Frame rate and container
30 fps, MP4
30 fps, MP4
30 fps, MP4
Audio wording in the features column
audio, audio toggle
audio sync
audio-driven
Smart duration
listed
not listed
not listed
Adaptive aspect ratio
listed
not listed
not listed
First-and-last-frame wording
covered by the all-in-one reference entry
not listed
listed as first frame and first/last frame
LIVE PRICING
How Wan 3.0 is billed
Alibaba Cloud bills wan3.0-video by the second, and the seconds it counts are not only the ones it produced. The documented rule is that the billed duration is the input video duration plus the output video duration: a reference video you supply is charged for the seconds it actually runs, and the clip that comes back is charged for the seconds it actually generated. A request that fails is not billed and does not consume the trial allowance. Turning audio off does not change the price, because the audio track is generated in the same pass rather than bought separately. The list prices read from the official pricing page on 15 September 2026 are 0.05, 0.10 and 0.20 US dollars per second for 480P, 720P and 1080P in the Singapore region, with 0.068, 0.14 and 0.28 for the speed-optimised variant. The same page lists a limited-time discount on the standard model and states no start or end date for it, so this page quotes the list price only and does not republish a discounted figure. The Chinese and English versions of that pricing page also disagree about which region holds the trial allowance, one naming Singapore and the other naming Beijing, so no trial allowance is described here either. This site does not bill in dollars; the workspace quotes platform points, and the live estimate before submission is the price that applies here.
The workspace quotes points before you submit. Points are set by this site and are not the official per-second figures, and the two systems are not convertible.
02
Resolution, duration and audio on this route
The model outputs 480P, 720P or 1080P, defaults to 1080P, runs from 2 to 30 seconds, and produces audio unless it is switched off. This page exposes 720p and 1080p from 2 to 15 seconds and has no audio switch, so it omits the official default resolution, omits the 480P tier, stops 15 seconds short of the official ceiling, and cannot turn the audio track off.
06 / Workflow
Running one generation end to end
01
Decide which of the three routes you are on
Text-to-video starts from a prompt with nothing uploaded. Image-to-video starts from one uploaded frame. Reference-to-video starts from one uploaded reference video. The three cannot be combined in a single submission, so choosing the route is the first decision rather than an option set later in the form.
02
Write the movement rather than the mood
The model reads a description of what changes, so naming the camera behaviour, the direction of travel, the speed and what must stay still is more useful than adjectives about atmosphere. Saying that the camera does not move, or that a label must stay legible, is a working instruction rather than a preference.
03
Keep the duration inside what this page offers
The selector here runs from 2 to 15 seconds with 5 as the default, while the model itself accepts 2 to 30. Five seconds is the documented default on both sides and is the safest place to start, because it matches the length the model reaches for when no duration is given at all.
04
Pick the resolution knowing which tier is the official default
The official default is 1080P, and this page defaults to 720p while offering both 720p and 1080p. If the deliverable has to match what an official call would have produced, set the resolution explicitly rather than leaving the standing default in place.
05
Read the estimate, then review the whole clip
The workspace shows a point estimate before submission, which is the only price that applies here because this site bills in points rather than seconds. After generation, watch the clip from end to end rather than the first two seconds: identity drift, shifting geometry and text that stops being legible tend to appear after the point most people stop watching.
Review before publishing
Ceilings, constraints, and what this page does not offer
The usual failure modes apply. Faces drift across a clip, geometry shifts, a subject changes shape partway through, and text that reads cleanly in a still frame turns illegible once it moves. The thirty second ceiling and the multi-shot mode both raise the exposure, because there is more time for a drift to become visible and because a cut inside a generation is another place for continuity to break. Generate at the shortest length your judgement permits, watch the whole thing rather than its opening, and keep the prompts that worked, because the model reads a movement description literally and the same wording tends to produce the same result.
The distance between the official model and this page is wide enough that it should be read as two different surfaces. The official side accepts ten reference images, five reference videos, five reference audio clips, a document or a public link, a first frame and a last frame, and it can edit or extend a supplied clip. This page accepts one reference video or one frame or a prompt, and nothing else. The two exceptions run the other way: the reference video accept list here includes WebM, which the official format table does not list, and the standing default resolution here is 720p where the official default is 1080P. Neither exception makes up for the missing inputs, and no page should describe the model as if the workspace were the model.
One constraint inside the model itself is worth stating separately because it inverts a natural assumption. Where the Wan 2.7 image-to-video series accepts driving audio together with a first frame or a first-and-last-frame pair, the Wan 3.0 first-frame and first-and-last-frame modes do not. The documentation directs anyone who needs audio-driven generation to the all-modal reference mode instead, combining a reference image with a reference audio clip. A newer version number is therefore not a superset of the older one, and treating it as one produces a request the API rejects.
Two boundaries sit outside the model entirely. The first is that the pricing page publishes a limited-time discount on the standard model without giving it a start or an end date, and the same page states in one language that the trial allowance is available only in Singapore while the other language version says only in Beijing. This page therefore quotes the list price and describes no trial allowance, because a figure whose window and region are both unstated cannot be published responsibly. The second is that the task id and the result URL both expire 24 hours after the task succeeds, so anything worth keeping has to be transferred somewhere else rather than linked to later.
Sources
Official sources and how these facts were checked
Specifications, input limits, billing rules, rate limits and lifecycle status come from Alibaba Cloud Model Studio documentation and from the official community blog post, read at the source rather than through third-party summaries. The documentation is served without a login and none of it is behind a paywall. Platform controls and points come from this site. Where the two differ, the difference is written out in the comparison above instead of being smoothed over.
They are the display name and the identifier. The documentation spells the model Wanxiang 3.0 and writes it as Wan3.0, while the string a request has to carry is wan3.0-video, with a dot inside the version number and a -video suffix. The speed-optimised variant is wan3.0-video-prime. The route on this site is named wan-3-0 with hyphens, which is this platform naming rather than a model id.
What does All-in-One actually mean here?＋
The documentation says the model covers all of its task types without the caller switching model names, and that it routes automatically from the type field in the media array together with the intent in the prompt. Text-to-video, first-frame, first-and-last-frame, multi-modal reference, document or link input, editing and extension are all the same id rather than separate endpoints.
How long can one generation be?＋
Between 2 and 30 integer seconds, with 5 as the documented default and -1 as a smart duration that lets the model settle on a length. When a video is supplied as input there is an extra rule: the input video duration plus the output video duration must not exceed 30 seconds, so a long reference clip reduces how much can be generated alongside it.
What resolution does it output, and which is the default?＋
480P, 720P and 1080P, and the documented default is 1080P rather than the cheapest tier. Aspect ratio accepts 16:9, 4:3, 1:1, 3:4 and 9:16, with adaptive as the default so the output follows the proportions of the source. Output is 30 frames per second in an MP4 container, and one request returns one video.
Is the audio generated or added afterwards?＋
Generated in the same pass as the video. The documentation describes dialogue, sound effects and background music coming out of the model directly, with no separate dubbing step. Audio is enabled by default and can be switched off with the audio parameter set to false, which removes the track without changing what the request costs.
How is a request priced?＋
By the second, and the seconds counted include the ones supplied as well as the ones produced. The documented formula is that the billed duration equals the input video duration plus the output video duration. Failed requests are not billed. List prices read from the official pricing page on 15 September 2026 are 0.05, 0.10 and 0.20 US dollars per second for 480P, 720P and 1080P in the Singapore region.
Can I get a discount, and is there a free allowance?＋
The official pricing page marks a limited-time discount on the standard model and gives no start or end date for it, so this page quotes the list price and does not republish a discounted figure. The same page disagrees with itself about the trial allowance, naming Singapore in English and Beijing in Chinese, and lists an allowance of 30 seconds valid for 90 days either way. Because the region is unresolved in the official sources, no allowance is described here.
Does the first-and-last-frame mode support audio-driven generation?＋
No, and this is a documented regression rather than an oversight. Where the Wan 2.7 image-to-video series accepts driving audio alongside a first frame or a first-and-last-frame pair, the Wan 3.0 first-frame and first-and-last-frame modes do not. Anyone who needs audio-driven generation is directed to the all-modal reference mode, combining a reference image with a reference audio clip.
What can I upload, and in what quantities?＋
Up to 10 reference images at 20 MB each, up to 5 reference videos totalling 15 seconds at 100 MB per clip, up to 5 reference audio clips totalling 15 seconds at 15 MB per clip, and one document of up to 50 pages or one public page that needs no login. Reference media, a document or a link cannot be combined with a first frame or a first-and-last-frame pair; the API rejects that mix with an InvalidParameter error.
What does this page actually offer?＋
Text-to-video, image-to-video from one frame, and reference-to-video from one reference video, with duration from 2 to 15 seconds, 720p and 1080p output, and five fixed aspect ratios. There is no first-and-last-frame route, no reference image upload, no audio input or audio switch, no document or link input, and no editing or extension mode.
Does this page call the official API?＋
No claim is made either way. This is an independent creation platform that runs its own generation route, and it is not affiliated with Alibaba Cloud or the Tongyi Wanxiang team. The official model does have a documented API, but the route behind this page is this platform's own route, and the page does not assert an integration it cannot evidence.
How long is a result available after it is generated?＋
The task id and the output URL both expire 24 hours after the task succeeds, and querying a task after that returns an unknown status. Anything worth keeping has to be moved somewhere else inside that window rather than linked to afterwards. Rate limits are 5 requests per second and 5 concurrent tasks in every region the documentation lists.
Return to the live workspace, choose one of the three routes this page supports, keep the duration at five seconds while you are testing the description, and read the current point estimate before you generate.