Skip to content
Grok image
English
English简体中文繁體中文DeutschEspañolPortuguês

Wan 3.0

Wan 3.0 is the video model Alibaba Cloud sells as wan3.0-video through Model Studio and runs on wan.video. This page puts the publisher's own parameter reference next to the settings this site's service returns for the same product, and then next to what the workspace actually lets you attach, because on this model those are three different numbers.

Updated:

Wan 3.0

Choose a mode, write the prompt, then read the quoted cost before you generate.

Wan 3.0
AI Video0
5s
2s15s

Generated content will appear here

Source material

Official material for Wan 3.0

Frames from the publisher's own consumer site and its cloud scene page, which are the only two surfaces that serve generated output for this model. Four of the clips those frames belong to are published further down this page.

The developer documentation for this model ships no output media at all, so every frame here came from the consumer or cloud scene pages instead. No prompt is published by the publisher for any of these clips, so no prompt is attributed to an item, and the frame rate of these web files is not presented as the model's output rate.

A woman in a long red dress standing on a wet floor in front of a painted backdrop, with a film crew and equipment visible along the left edge of the frame Read the prompt ↗

01 / Frame from an official consumer-site reel

A costume drama staged on a built set

WAN consumer site ↗

Prompt published with this item

The publisher does not publish a prompt for this clip

Open this prompt in the workspace ↗

The publisher does not publish a prompt for any of its Wan 3.0 clips, so no prompt is attributed to an item here. Where text appears inside a picture, it is part of the scene the publisher published, not a caption this site added.

Two animated fighters in flat bold outlines, one with blue hair and white straps and the other in a yellow cap and green jacket, trading a blow against a red background Read the prompt ↗

02 / Frame from an official consumer-site reel

A drawn fight, kept on model across the cut

WAN consumer site ↗

Prompt published with this item

The publisher does not publish a prompt for this clip

Open this prompt in the workspace ↗

The publisher does not publish a prompt for any of its Wan 3.0 clips, so no prompt is attributed to an item here. Where text appears inside a picture, it is part of the scene the publisher published, not a caption this site added.

A blue plush creature with a very wide open mouth full of flat teeth and a small plush toy held in it, photographed like a soft-toy product against a starry background Read the prompt ↗

03 / Frame from an official cloud scene page clip

Plush material in a plush universe

Alibaba Cloud Wan3.0 scene page ↗

Prompt published with this item

The publisher does not publish a prompt for this clip

Open this prompt in the workspace ↗

The publisher does not publish a prompt for any of its Wan 3.0 clips, so no prompt is attributed to an item here. Where text appears inside a picture, it is part of the scene the publisher published, not a caption this site added.

A black compact camera with a silver lens ring photographed square-on against a plain off-white background, casting a soft shadow Read the prompt ↗

04 / Frame from an official cloud scene page clip

A product surface rendered cleanly on a plain background

Alibaba Cloud Wan3.0 scene page ↗

Prompt published with this item

The publisher does not publish a prompt for this clip

Open this prompt in the workspace ↗

The publisher does not publish a prompt for any of its Wan 3.0 clips, so no prompt is attributed to an item here. Where text appears inside a picture, it is part of the scene the publisher published, not a caption this site added.

What Wan 3.0 is, in the publisher's own words

Alibaba Cloud presents this model as a single endpoint that covers several ways of making video. The Chinese API reference opens with the sentence "万相 3.0 是全能参考视频生成模型(All-in-One),统一支持 文生视频 、 图生视频 (首帧/首尾帧)和 参考生视频 等多种用法", which is that Wan 3.0 is an all-in-one reference video model covering text to video, image to video with a first frame or with a first and last frame, and reference to video. The English guide states the same envelope in numbers: "supports up to 30 seconds per generation, 30fps output frame rate, natively outputs dialogue, BGM, and sound effects, supports up to 20 multi-modal reference materials (images, videos, audio, documents, web pages) per request, and supports first frame/first-last frame control, video editing and extension".

The scene page on Alibaba Cloud's own site makes the reference claim more concrete: "支持文本、图像、音频、视频四大模态输入,最高支持 20 个素材参考生成" — four input modalities, up to twenty reference materials. The consumer site at wan.video lists the same claims in its own banner: native 30 seconds, up to 20 reference assets with document and webpage parsing, pixel-perfect consistency, and precision video editing.

Everything in this section is quoted from the publisher. Nothing on this page is a statement by Alibaba Cloud, and this page has not been reviewed or endorsed by them.

Quotations are left in the publisher's original language, because translating a specification is one of the ways numbers drift.

One endpoint, seven task types, and no switching between them

This is the first thing to get right, because the service presents the same model differently. The publisher ships one model name and routes internally. The guide says so directly: the model "covers all the following task types without switching model names", and it routes "based on the type field in input.media and the prompt intent". The task table in that same guide lists seven task types under that one name.

The routing is not a convenience wrapper. The reference notes that "type values of different task types cannot be mixed", so a first-and-last-frame request accepts only first_frame and last_frame and cannot carry a reference audio clip at the same time. Which task you are doing is decided by what you attach, and the attachments themselves decide what else is allowed.

This service takes the opposite approach and re-splits the endpoint. Its entry for this product declares three modes — text to video, image to video and reference to video — with three separate task names behind them. Two of the publisher's seven task types survive that split as first-class modes; one, first and last frame, is absorbed into image to video as a second image; and four are not declared at all, which is where file and web page reference, video editing and video extension sit.

It is worth being precise about what that costs. Someone arriving from a search for Wan 3.0 video editing will not find an editing form here, because this product does not declare one, not because the model cannot do it.

Task type the publisher documentsHow it is triggeredLimits stated
Text to videoprompt only, no mediaresolution, aspect ratio and duration all free
First frame to videotype set to first_frameaspect ratio recommended as adaptive
First and last frame to videotypes first_frame and last_frame togetherone image each, strictly the first and last frame
Multi-modal referencereference_image, reference_video or reference_audio, in any combinationup to 20 materials: 10 images, 5 clips, 5 audio
File or web page referencetype file or link, one of the twoone each, and the two cannot be combined
Video editinga reference video plus an editing instruction in the promptratio adaptive, duration -1 recommended
Video extensiona reference video plus an extension instructionforward, backward or both directions

What this site's service publishes for the same product

The workspace does not read the publisher's documentation. It reads this site's own service catalogue, which answers with one entry per model. The entry for this page is named Wan 3.0, sits in the family the catalogue calls Wan, and is marked available and callable. It declares three modes and carries a published floor of 6 points per second, with the billing mode labelled parameters plus duration.

A floor is not a quote. The server totals the charge in points for the mode and settings you choose before a task starts, and that total is the one that counts.

Mode the service declaresTask name behind itFields in the published form
text-to-videowan3.0-video-text-to-videoprompt, resolution, aspect ratio, duration
image-to-videowan3.0-video-image-to-videothe four above plus reference images, 1 to 2
reference-to-videowan3.0-video-reference-to-videothe four above plus reference images 0 to 10 and reference videos 0 to 5

Every parameter the service publishes, set against the publisher's table

The publisher's parameter reference gives this endpoint a small set of fields, and the service publishes most of them. Several agree outright, and two do not.

The duration agreement is exact, and it is stated twice by the publisher rather than once. The parameter table gives the range, and the guides print the same thing in a summary line: duration equals 2 to 30 seconds or minus one for smart duration.

The publisher also publishes a second duration rule that has no counterpart in the form at all. When a video is attached, the input and the output together are capped at 30 seconds. That cap does show up in how the mode is priced, which is the subject of a later section.

There is also a third source of settings in this project, and it is more wrong here than on any page this site has built before. This site's source code carries a local family table used only when the service publishes no value for a field. For this family it declares a duration range of 2 to 15, quality tiers of 720p and 1080p with no 480p, five aspect ratios with neither adaptive nor 21:9 and a default of 16:9, audio off by default, and a fallback cost of 80. The publisher and the service both say 30 seconds, both say 480p exists, and both default the aspect ratio to adaptive. None of those local values is consulted, because the service publishes all four of those fields in its form.

The local table is described here because one of its values is not inert, and that one is the reason this page exists. It is covered two sections down.

SettingPublisher's API referenceThis service's published form
Durationdefault 5, integers from 2 to 30default 5, minimum 2, maximum 30
Duration with a video attachedinput duration plus output duration must not exceed 30 secondsnot published as a field; the basis of the charge changes instead, see below
Aspect ratioadaptive by default, with adaptive, 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16adaptive by default, with adaptive, 16:9, 4:3, 1:1, 3:4 and 9:16
Resolution tiers1080P, 720P and 480P480p, 720p and 1080p
Resolution default1080P720p
Audioa boolean, true by default, and switching it costs the same either wayno field in the form
Seedminus one, or 0 to 2147483647no field

Two settings the service publishes differently: the resolution default, and the wide frame

These are the two places where a choice made here produces a different result from the same choice made on the publisher's own platform, and both are visible before you submit anything.

The resolution default is worth stating in money rather than in pixels, because the price list published for this product is keyed on resolution as well as on reference material. For the class with no reference material at all, the published rates are 6 points per second at 480p, 12 at 720p and 23 at 1080p. A five-second clip is therefore 60 points on the default here and 115 points at the publisher's default.

The aspect ratio is the narrower of the two gaps and also the easier one to misread. Alibaba Cloud's guides print the ratio list as 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, and its API reference prints the same six with adaptive added in front of them and named as the default. The service publishes six values too, but its six are adaptive plus five, and 21:9 is the one it leaves out. The publisher's own worked example for text to video opens with a prompt describing "20 秒超宽银幕" material, so the wide format is not a footnote in their material.

Choosing adaptive is not the same as choosing 21:9. Adaptive means the model picks a ratio from the input media and the intent, and a text-only request has no input media to read. So a fixed ultra-wide output for a text-to-video job cannot be requested from this workspace, and this page does not claim that it can be reached by another route.

SettingPublisher's defaultDefault hereWhat the difference means in practice
Resolution1080P720pthe first generation on this page costs roughly half of what the same settings would cost at 1080p
Fixed 21:9 outputselectablenot selectableadaptive ratio can follow a wide input, but there is no way to pin the output to 21:9 from text alone

How a reference request is priced, and what the price list prices that the form cannot send

This product publishes a price list rather than a single rate, and it is the most detailed one on this site. Fifteen rows cover five classes of reference material across three resolutions. Two things about it are worth reading carefully before they surprise anyone.

Nothing in this table is presented as a total. The server returns the total for your settings before a task starts, and that is the only figure that binds.

  • The audio row prices a class this form cannot produce. Reference to video publishes two attachment fields, images and videos, and no audio field anywhere. A rate exists for a request that cannot be assembled here.
  • The same fifteen rows are attached to all three modes. Text to video has no attachment fields at all, and it is still presented with rates for four classes of reference material it cannot accept.
  • Attaching a video changes the basis of the charge, not only its rate. The service's own routing table gives text to video and image to video a basis of output seconds, and gives reference to video a basis of input plus output seconds. The pricing block says the same in words: video reference is charged as the larger of the input video's length and the output length, plus the output length.
  • That last rule is the one place where the service echoes the publisher exactly. Alibaba Cloud caps a request with video input so that the input duration plus the output duration does not exceed 30 seconds, which is what makes an input-plus-output basis possible at all.
Reference material attached480p720p1080p
none6 points per second1223
images61223
audio61223
video122345
mixed122345

Reference assets: twenty published, fifteen selectable, fewer reachable

Three numbers describe the same capability, and they are not the same number. The publisher's headline is twenty multi-modal materials per request. The service's form can accept fifteen of those. The workspace can send fewer than that, and the reason sits in this site's own code rather than in anything the service publishes.

The result is a ladder rather than a single gap: twenty documented, fifteen accepted by the form, one reachable from the interface. Every step of that ladder is checkable, and the last step is the one nobody would predict from reading either the publisher or the catalogue.

There is a wrinkle worth recording because it shows the two mechanisms colliding inside one product. The audio toggle on this page is decided by the service's form, following the rule that the workspace renders an audio control only when the form carries the field for it. The reference-image limit on the same screen is decided by the local table instead. So two controls on one form are answering to two different sources, and for this product they disagree.

Nothing about this is a claim that the model is weaker than its documentation. The generation is whole; what is missing is the path to ask for it. This page states the limits and leaves the routing question where it belongs.

  1. The publisher documents twenty reference materials in one request: up to 10 images at 20 MB each, up to 5 video clips at 100 MB each and 15 seconds in total, and up to 5 audio clips at 15 MB each and 15 seconds in total, with the three kinds combinable in any order.
  2. The service's reference-to-video form publishes image_urls with a maximum of 10 and video_urls with a maximum of 5. That is fifteen, and there is no audio field, so the five audio materials the publisher allows are declared in the catalogue's capability block and cannot be submitted.
  3. The workspace renders its reference-image uploader only for the image-to-video and start-and-end modes. On the reference-to-video form it renders no image uploader at all, even though the service publishes a field accepting ten images.
  4. Where the workspace does render the image uploader, it clamps the selection to a number taken from the local fallback table rather than from the service. For this family that number is one for image to video and zero for reference to video. So image to video accepts a single image, and the publisher's first-and-last-frame method has no way to be expressed.
  5. The reference-video input accepts one file. The publisher allows five clips, and the service publishes a maximum of five.

Audio is declared, defaulted on, and not switchable here

The publisher treats audio as on and gives it a switch. This service declares audio as generated, defaulting to on, with a toggle. The form this page actually reads has no audio field, so no switch appears — which means the default is the only behaviour available.

Three separate facts are true at once, and it is worth keeping them apart. The catalogue's capability block for this product says audio is supported, is generated, defaults to true, and has a toggle. The publisher's parameter reference gives audio a default of true as well, and adds that turning it on or off costs the same either way, so there is no price reason to prefer one. And the reference-to-video row still declares an audio reference limit of five clips, which is the same declaration as the capability block.

None of that reaches the screen. The fields published for this mode are prompt, resolution, aspect ratio, duration, images and videos. The workspace builds its controls from that form, and it draws an audio switch only when the form contains a field for one.

The honest description is not that sound is off. It is that sound is on and cannot be turned off here, and that the reference-audio half of the same capability has no way in.

Publisher's audio surfaceWhat it doesPresent in this service's form
audioinclude sound in the output, true by default, same price either wayno
reference_audiouse up to five audio clips, fifteen seconds in total, as referenceno field, though the capability block declares the five-clip limit

What other pages for this term get wrong

Five pages were measured for this one. Most disagreements with the publisher are ordinary drift, and one is more interesting than the rest because it is not a wrong number at all.

The last row is the one to leave alone rather than settle. A note on Alibaba Cloud's own developer site says the parameter was deprecated for this model and must not be used, while the official API reference on the same company's documentation domain still carries it in the request bodies it tells developers to send. This page records the contradiction and does not resolve it, because nothing here can.

The 4K claim is the one that would change a decision, and it is a single line in a reference work rather than a measurement. The publisher's own parameter table lists three tiers and stops at 1080P.

There is also something about this page set that is more informative than any single claim. Four of the five pages carry no video element at all, and the one that does is an encyclopedia entry. One of them is worse than empty: the model page of a third-party API reseller is marked index, follow with eleven language alternates, and renders thirty-four words of body text under a second-level heading that reads "Something went wrong!".

ClaimWhere it appearsWhat the publisher documents
A gateway model string of alibaba/wan-3.0a third-party pricing write-upthe publisher's own model names are wan3.0-video and wan3.0-video-prime
Reference clips and audio clips each totalling 15 secondsthe same write-up, reading the publisher's limits correctlyup to 5 clips totalling 15 seconds and 5 audio clips totalling 15 seconds
Native 4K resolutionan encyclopedia entry480P, 720P and 1080P, with 1080P as the default
A per-second price stated only as a discounted figuredocuments published on the cloud provider's own community sitethe list price is per second and the discount window is stated with an end date
The prompt_extend parameter being deprecateda community article on the cloud provider's own developer sitethe official API reference still shows prompt_extend set to true in its own request examples

What this gallery shows, and where it came from

The publisher's developer documentation for this model carries no generated output at all. Across all six archived first-party pages there is not one video element and not one clip file; the pages are parameter tables and guides, and what illustrations they carry are interface screenshots. The model's output lives on the two consumer surfaces instead, and both of them had to be rendered rather than fetched, because both are assembled in the browser.

Ten clips were retrieved from those two surfaces and every one of them downloaded at full size. Four are published here as clips, chosen because their length or their movement is itself the evidence. A thirty-second clip is the only item in the set that reaches the length the publisher's headline claims, and it is published as a clip for exactly that reason. A five-second clip sits at the published default duration and is short enough to publish whole. An eight-second clip cuts between two contrasting setups of one scene, so shot switching is visible rather than described. A fifteen-second clip in 1080p carries the register the publisher calls digital information, where product surfaces and clean interfaces are the subject.

Six further clips were kept back and sampled into stills instead: two long reels from the consumer site's front page, at roughly twenty-seven and forty-four seconds, plus clips in distinctly different registers. Sampling is the honest treatment for a reel, because a still of a montage is a still of one moment in it, and the captions say which moment.

One frame-rate fact has to be stated carefully to avoid a wrong inference. The publisher documents a 30 frames per second output, and the ten clips retrieved run at 24, 25 and 30. Those are web deliverables rather than model output, so the frame rate is a property of these files and is not used anywhere on this page as evidence about the model.

Nothing here is attributed to an individual mode. The publisher does not publish a prompt or a parameter line beside any of these clips, so no caption claims one was produced from a particular reference material, a particular resolution or a particular duration. Where a caption mentions that a frame belongs to a clip of a given length, that length was measured from the file.

How to read the clips and stills on this page

The material on this page is the publisher's, and the labels say so on every item.

Where a comparison appears on this page, both halves come from the same clip, so the difference shown is a change inside one generation rather than a difference between two of them.

  • Uploaded clips are included rather than linked, and each caption says what was done to them: the video bitrate was reduced and the resolution, frame rate and audio track were retained.
  • Stills are frames sampled by this site from published clips, and each caption gives the point in the clip it was taken from. They are not separate images the publisher released.
  • No prompt is published for any of these clips. Every gallery item says so, rather than carrying a prompt this site invented or borrowed from a different model's example.
  • The four clips are published because of their duration or their movement, and the stills exist to show registers the clips could not carry at a viewable size.

Examples

Thirty seconds, published by the publisher on its consumer site and kept whole here because length is the claim being examined. This is the only clip retrieved for this page that runs to the thirty seconds the model's headline mentions, and it is a single continuous scene with dialogue rather than a montage. The audio track is present in the transcoded file.

WAN consumer site, retrieved 2026-09-19 from cdn.wanxai.com. Measured at 30.05 seconds, 1280 by 720, 24 frames per second. Transcoded from 1.21 Mbps to 1.40 Mbps; resolution, frame rate and audio retained. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Eight seconds in which one scene is shot from two contrasting setups, published as the demonstration of consistency under a cut. Two figures in bulky armour lit entirely by blue practical light are kept consistent across the change of distance, which is the property the publisher's page claims and which a still cannot carry.

WAN consumer site, retrieved 2026-09-19 from cdn.wanxai.com. Measured at 8.02 seconds, 1280 by 720, 25 frames per second. Transcoded from 3.34 Mbps to 1.80 Mbps; resolution, frame rate and audio retained. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Five seconds at the published default duration, published whole because water, cloth and animals in one frame is a hard combination and five seconds is short enough to publish without cutting it down.

WAN consumer site, retrieved 2026-09-19 from cdn.wanxai.com. Measured at 5.02 seconds, 1280 by 720, 30 frames per second. Transcoded from 2.61 Mbps to 1.62 Mbps; resolution, frame rate and audio retained. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Fifteen seconds at 1920 by 1080, the highest resolution in this set. Published as a clip because its subject is surfaces rather than people: product forms, a clean background and an object held square to the camera. Rendered output that reads as a product photograph is a different test from cinematic footage, and this is the only clip here that runs it.

Alibaba Cloud Wan3.0 scene page, retrieved 2026-09-19 from bailian-bmp-pre.oss-cn-hangzhou.aliyuncs.com. Measured at 15.00 seconds, 1920 by 1080, 30 frames per second. Transcoded from 4.94 Mbps to 2.00 Mbps; resolution, frame rate and audio retained. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Before

Before: On the left two armoured figures seen at a distance in a wide blue-lit space, one striking and one recoiling; on the right the same two figures in a tight framing, lit blue and white, one gripping a glowing blade

After

After: On the left two armoured figures seen at a distance in a wide blue-lit space, one striking and one recoiling; on the right the same two figures in a tight framing, lit blue and white, one gripping a glowing blade

The opening and closing setups of the same eight-second clip, side by side, so the change of framing can be read without playing the file. Both figures keep their armour detail, their proportions and the direction they face across the change of distance, which is what the consistency claim is about.

Frames sampled by this site at 2.0 and 6.8 seconds from the publisher's eight-second clip, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A film set seen from the side, with crew members, a boom pole and lighting stands around a woman in a red dress walking away from the camera

Sampled at 6.7 seconds from the longer of the two consumer-site reels. A behind-the-scenes register, staged as if filmed on a working set.

Frame sampled by this site from the publisher's consumer-site reel, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A performer in a red dress seen from behind in a warm-lit theatre, with a second figure mid-gesture on the left and red curtains behind

Sampled at 22.9 seconds from the same reel. The register changes inside one generation, from set documentary to stage performance.

Frame sampled by this site from the publisher's consumer-site reel, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A man in a dark jacket and open-necked shirt aiming a handgun toward the camera, his face wet, against a dark background

Sampled at 24.1 seconds from the second consumer-site reel. Low-key lighting with the face still readable, which is the harder half of that look.

Frame sampled by this site from the publisher's consumer-site reel, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A person seated at a desk in a sunlit room with patterned wallpaper and a bicycle, a cat standing on the desk beside them, seen from behind

Sampled at 37.2 seconds from the same reel. A domestic interior where light through the window and the detail on the desk have to hold together.

Frame sampled by this site from the publisher's consumer-site reel, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A stylised young figure with long pale hair sitting on a kitchen floor eating from a tub, with an open fridge behind showing shelves of food

Sampled at 17.0 seconds from a twenty-second consumer-site clip. A rendered register rather than a photographic one, with readable packaging and dense product detail inside the fridge.

Frame sampled by this site from the publisher's consumer-site clip, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A musician in dark clothing playing a violin in heavy rain, seen in profile against a deep teal background

Sampled at 3.9 seconds from a seven-second clip. Published for the hands: the bow, the strings and the fingers stay legible at close range while rain crosses the whole frame.

Frame sampled by this site from the publisher's consumer-site clip, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A woman in a long crimson cloak walking through shallow water, with several dark horses moving behind her in mist

Sampled at 1.2 seconds from the five-second clip. Costume, water and animals together in one frame, which is where cloth and animal shape most often drift.

Frame sampled by this site from the publisher's consumer-site clip, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A performer in a red robe sitting on sand dunes at sunset playing a long-necked stringed instrument, desert and clouds behind

Sampled at 17.0 seconds from a twenty-second clip on the cloud scene page. A warm sunset gradient across an entire frame, with a small object held and fingered correctly at the centre of it.

Frame sampled by this site from the Alibaba Cloud Wan3.0 scene page clip, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

A black compact camera photographed square-on against a plain off-white background, its lens ring and dials sharply defined

Sampled at 12.8 seconds from the fifteen-second 1080p clip. A product form on a seamless background, which is the surface the publisher's own page singles out as a target for this model.

Frame sampled by this site from the Alibaba Cloud Wan3.0 scene page clip, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.

Questions

Is this the official Wan 3.0 page?

No. It is a page on this site about the model. The publisher's own pages are named as sources throughout, and this page has not been reviewed or endorsed by Alibaba Cloud.

Which modes can I run here?

Three: text to video, image to video and reference to video. The publisher documents seven task types for the same model name, including file and web page reference, video editing and video extension, and this product does not declare any of those three.

Why does the publisher call it one model and this page three modes?

Because the publisher routes internally and this service does not. Alibaba Cloud ships one model name and picks the task from what you attach and what your prompt asks for. The catalogue here splits the same endpoint into three selectable modes, which is a different shape for the same capability.

How many reference images and videos can I attach?

The publisher documents up to 10 images and 5 clips, plus 5 audio clips. The service's form accepts 10 images and 5 videos and has no audio field. The workspace is the narrow end: its reference-video input takes one file, and its reference-to-video form renders no image uploader at all.

Can I make a first-and-last-frame video?

Not from this workspace. The publisher documents it as its own task type with one image for the first frame and one for the last. The service's image-to-video form accepts up to two images but does not distinguish which is which, and the workspace clamps that upload to a single image.

Can I ask for a fixed 21:9 output?

No. The publisher's ratio list includes 21:9 and this service's list does not. Choosing adaptive lets the model match an input's proportions, but a text-only request has no input to match, so a fixed ultra-wide frame cannot be requested here.

The publisher defaults to 1080P. What does this page default to?

720p. The publisher's API reference names 1080P as the default for this endpoint and this service's form names 720p. The difference matters because the rates published here are keyed on resolution as well as on reference material, and the default costs about half of what the publisher's default would.

Why does the rate table have an audio row when there is no audio field?

That is what the catalogue publishes. The rate table covers five classes of reference material across three resolutions, and one of those classes is audio, while the form for the same mode publishes only image and video fields. The same table is attached to text to video, which has no attachment fields at all.

Can I turn the sound off, or attach audio as a reference?

Neither, from this form. The service's capability block declares audio as generated and defaulting to on with a toggle, and the publisher defaults audio to true as well and charges the same either way. The published form has no audio field, and the workspace draws an audio control only when the form contains one.

What does it cost?

The server quotes it in points for the mode and settings you chose, before the task starts. The catalogue publishes a floor of 6 points per second for this product, and a rate table that varies with resolution and with what you attach. Attaching a video also changes the basis of the charge to input plus output seconds.

Explore next