HappyHorse 1.1
An Alibaba Cloud video model with a reference mode that takes up to nine images. The mode is published, priced and callable, and the workspace on this site will not open it. This page shows why, line by line, and names the origin of everything it displays.
Happy Horse 1.1
Choose a mode, write the prompt, then read the quoted cost before you generate.
This model does not support reference video input
Results
Generated content will appear here
What this page is, and the name it goes by
This is a page about a model, not a page by the model's publisher. The product is HappyHorse 1.1. Alibaba Cloud announced it on 22 June 2026 and also calls it 快乐小马, which is the name its Chinese consumer site carries. Two consumer domains belong to it: happyhorse.cn serves the Chinese platform and happyhorse.com the English one. The two cross-link each other and load the same build from the same Alibaba CDN path, which is how they are identified here, because the publisher's own product-news entry names neither domain.
The name is the first hazard. HappyHorse fuses an English adjective and an English noun, so a search for the spaced form returns dictionary entries rather than the model, and at least eight third-party domains have been built on variants of it without any of them being the publisher. Several publish confident capability claims. One publishes a nine-value aspect-ratio matrix with no source. None of them is used as a source on this page.
The second hazard is closer to home. The catalogue this site reads names just one source address for the product, and that address points at a reseller page rather than at Alibaba. Anything this page states about the publisher comes from an Alibaba Cloud or Qianwen Cloud page that has been archived and checked against its own title, and anything it states about a reseller is attributed to that reseller by name.
The publisher's product-news entry, the Qianwen Cloud model page for this product, both consumer homes and one Alibaba developer-community article are archived under docs/qa/official/. Each one was verified by its own page title before use, because two surfaces in this project have previously turned out to be not-found shells that answered with a 200 status.
Three modes this service publishes for one model
The service catalogue returns three task types under the single product name, and marks all three available and callable. Each one carries its own form, so the fields a visitor sees change with the mode tab they choose.
Two details in that table are worth reading twice. The image-to-video form carries no aspect-ratio field at all, while the other two carry one. And the reference mode accepts up to nine images, which is the largest reference budget anywhere in this catalogue.
All three modes bill on output seconds. That is unusual in this catalogue, where the reference mode of some other models bills on input plus output, and it means attaching nine reference images here does not change the basis of the charge.
| Mode | Model key the catalogue returns | Form fields beyond the prompt | Inputs the catalogue accepts |
|---|---|---|---|
| Text to video | happyhorse-1.1-t2v | resolution, aspect ratio, duration | none |
| Image to video | happyhorse-1.1-i2v | resolution, duration, one image | exactly 1 image |
| Reference to video | happyhorse-1.1-r2v | resolution, aspect ratio, duration, images | 1 to 9 images |
The reference mode is published, and the workspace will not open it
On this site a visitor cannot choose reference-to-video for this model. The tab is present and disabled, with a tooltip saying video upload is not supported. The reason is not a decision about HappyHorse; it is a gate written for a different shape of model, and it has been read out of the source.
The consequence is a contraction in three steps. The publisher documents and promotes nine reference images. The catalogue accepts nine. The number a visitor can attach is zero, because the mode that carries the image field is the mode the interface refuses to open.
The local fallback table inside this repository also says this family does not do reference-to-video, so the visible outcome would be the same if the catalogue were unreachable. The two agree by accident. The fallback says the model has no such mode; the gate says the mode must take a video. Both statements are in the source and they are not the same statement.
This is written from the shipped source of the workspace, not from a running session, and it is described as the code states it. Earlier pages in this batch were built on the same reading and the browser acceptance pass confirmed the disabled state in every locale.
- The workspace builds the list of models offered for a mode. For reference-to-video it filters that list so a model qualifies only if its reference form contains a field of type video, or a field named video, videos or video_urls.
- HappyHorse's reference form contains prompt, resolution, aspect ratio, duration and an image field. There is no video field on it.
- The same test decides whether the reference tab is enabled, so the tab renders disabled and the video input beside it renders disabled as well.
- The reference-image uploader is drawn only for image-to-video and start-end-to-video, so the reference mode would have had no image slot either.
What the publisher says nine reference images are for
The reference mode is not a side feature in the publisher's telling. Alibaba's product-news entry lists five upgrade dimensions over the previous generation and the community write-up expands each one; the reference budget is the headline of the consistency dimension.
The community article is not official documentation. It carries a copyright notice saying the content is contributed by a registered Alibaba Cloud user, and its register is promotional. It is cited here only for statements that a first-party page corroborates, and it is labelled as a community article wherever it appears.
What matters for this page is the mismatch rather than the claim. The capability the publisher promotes hardest for this generation is the one this workspace cannot reach, and nine is the largest published reference budget in the catalogue that the workspace then declines to use.
- The product-news entry, dated 22 June 2026, names dynamic expressiveness, character consistency, instruction following, visual texture and audio capability as the five dimensions upgraded over version 1.0.
- An Alibaba developer-community article describes the reference mode as the core commercial upgrade and states that up to nine reference images can be supplied at once to lock a character, an outfit, a product detail, a scene element or a brand mark.
- The same article frames the result as series content, where one character or product must survive across several shots and several clips.
- The Qianwen Cloud model page describes its image-to-video variant as improving skin texture, identity retention across clips, motion smoothness, text rendering stability and audio-video sync.
An aspect-ratio control on image-to-video that never reaches the request
The image-to-video form here shows five aspect ratios. Nothing a visitor picks there has any effect, and the interesting part is that this matches the upstream rather than contradicting it.
Here is the chain. The catalogue publishes no aspect-ratio field for image-to-video, so the workspace falls back to a table held in this repository, which supplies five ratios and draws the selector. When the request is assembled, only fields the selected mode declares are packed into it, and aspect ratio is not one of them. The value is dropped silently.
Two independent integrations say the same thing about the upstream. One documents the image-to-video endpoint as rejecting the aspect-ratio parameter and states that output framing follows the source image. Another lists aspect ratio as supported for text-to-video and reference-to-video only. The publisher's own worked example for image-to-video, archived from the Qianwen Cloud page, passes a model key, a prompt, a first-frame image, a resolution and a duration, and no aspect ratio at all.
A control that changes nothing is worse than an absent control, because it invites a visitor to make a choice the request will discard. It is recorded here rather than quietly hidden, and the fix belongs to the shared workspace rather than to this page.
| Source | Aspect ratio on image-to-video |
|---|---|
| Publisher worked example | not present |
| This service catalogue | field not published |
| Local fallback table in this repository | five values, selector drawn |
| Two independent integrations | documented as rejected upstream |
How many aspect ratios this product has
Three different answers are in circulation and none of them is obviously right, so all three are reported with their sources.
The fourth row matters more than it looks. EvoLink is the page this site's own catalogue points at as the source for the product, and its playground offers nine ratios under a selector that also defaults to 16:9. So the narrowing from nine to five happens inside this catalogue's own declared lineage, not between this site and a stranger.
No first-party page publishes a ratio list, so there is no tiebreaker. The page reports the three counts rather than choosing.
| Source | Count | Values |
|---|---|---|
| Publisher worked example | none shown | the example passes no aspect ratio at all |
| This service catalogue | 5 | 16:9, 9:16, 1:1, 4:3, 3:4 |
| EvoLink, the source address this catalogue names | 9 | adds 4:5, 5:4, 9:21 and 21:9 to the five above |
| Two further integrations | 9 | the same nine values |
The price this catalogue does not carry
For most products the catalogue publishes a points rule. For this one it publishes an exception fallback and nothing else, and the numbers that do appear match none of the published price lists.
The catalogue's points field for this product reads nineteen points per task and is labelled as an exception fallback, with a single detail row meaning that this applies when no rule matches. Its source identifier and pricing version are both empty. Two other video products in the same catalogue carry real rules instead: one publishes a floor of six points per second with fifteen detail rows, another a floor of five with three. So the absence is specific to this product rather than a general state of the catalogue.
This page therefore publishes no price of its own. A floor is not a quote in any case, and on this product there is not even a floor to quote. The charge is totalled by the server for the mode and settings chosen before a task starts, and that total is the only figure that binds.
The catalogue rate block's two numbers sit close to the official ratio between the tiers and close to no single published list. Converting between four currencies and four price lists would produce a number this page cannot defend, so it does not.
| Source | 720p | 1080p | Unit |
|---|---|---|---|
| Publisher list price | CNY 0.90 | CNY 1.20 | per second, with a 480P tier at CNY 0.45 |
| EvoLink, the source address this catalogue names | USD 0.124 | USD 0.160 | per second |
| Two further integrations | USD 0.14 | USD 0.18 | per second |
| This service catalogue rate block | 0.1364 | 0.1826 | per second, denomination not stated, no source recorded |
Audio, where five sources give five answers
Whether this model produces audio is the one capability question on which the sources disagree most, and the disagreement runs from the publisher's own news entry down to the audio track on a file.
This service does not answer the question either way. The form has no audio field, the capability block carries no audio key, and the workspace therefore draws no audio switch for this model. Whatever the model does with sound, a visitor here cannot ask for it or turn it off.
One further observation, kept strictly to what it is. The single playable file on this page carries an audio track. That is a property of the file. It is not evidence about the model, in the same way that the frame rate of a web deliverable is not the frame rate of a model.
| Source | What it says | Type |
|---|---|---|
| Alibaba Cloud product-news entry | audio capability is one of five upgraded dimensions | first party |
| Qianwen Cloud model page | audio-video sync listed among the improvements | first party |
| Alibaba developer-community article | a native audio matching system is described | community, not documentation |
| One integration | native audio is generated automatically | third party |
| EvoLink | the native audio row for this model is blank, while the same table credits another model with multilingual lip sync | third party |
Where every picture on this page came from
None of the pictures here was published by the model's maker, because the maker publishes no demonstration material at all on any surface that could be archived.
Four first-party pages were archived: the product-news entry, the Qianwen Cloud model page, and the two consumer homes. Between them they contain zero video elements and zero source elements. Their only images are platform chrome served from an Alibaba image host. The destination named by the news entry's try-it link is behind a console login, and its published route answers with a not-found page; the Alibaba benefits page carries fifty-seven images, all of them marketing art, and no video.
The consumer home is a creation feed rather than a showcase. Its items are served from a user-uploads path, each carries a named account and a Google account avatar, and not one of them labels the model that produced it. Publishing those works here would put a real person's unlabelled work under this model's name, so none of them is used.
What is left is the integration pages. Four stills come from the page this catalogue names as its source, one for each mode, and the platform's own alt text is what ties each picture to a mode. Two square images and one five-second clip come from a different integration's playground, where they are labelled as reference images and as the demonstration those references produce. That platform files the assets under the previous generation's name while serving them from the current model's page, so the attribution rests on the page that serves them; the captions say so.
Every item was screened image by image on two contact sheets. Rejected: three byte-identical copies of one platform's wordmark, three decorative cloud graphics that share a directory with that site's logo and footer badges, the model and provider logos across four sites, and the publisher's entire user feed.
Examples

The lead illustration on the page of the integration this site's catalogue names as its source, under that platform's own alt text for an AI video generation showcase. Wide enough to publish as a still rather than as a frame of something else.

Filed by the source platform under the text-to-video name and captioned by that platform as a text-to-video showcase, so the mode label is the platform's and not this site's inference.

The image-to-video illustration. Of the four it is the one that shows its mode most plainly, a row of stills feeding a stream of light, which is also the clearest visual counterpart to the inert aspect-ratio control described above.

The only picture in this set tied to the reference-to-video mode, tied by its filename. The source platform's alt text for this one names its own API rather than a mode, so the file and its alt text disagree; both are reported and the mode label is the filename's.
Five seconds of a fluffy pony loose in a commercial kitchen, pots flying and a fire breaking out. This is the only playable demonstration of the model found anywhere in the field, and it is the reference mode, the one this workspace disables. The prompt the platform publishes beside it asks for the first reference image to be placed inside the second while cooking disasters happen everywhere, which is what the clip shows.

Published as an input, not as output, because that is what it is. The platform loads this as a reference image for the clip above and labels it in its own alt text as a reference image. The pony is the subject the reference mode is meant to hold steady across a shot.

The second reference input the platform supplies for the same demonstration, likewise interpreted from the platform's alt text rather than by guessing. Pairing it with the pony is what makes the clip a reference-to-video example, since the two inputs carry both the subject and the setting.

A frame sampled one second into the five-second clip, before anything has gone wrong. Sampled here from the file rather than supplied as a poster by the platform, so the stills on this page are points of one continuous shot and not four separate pictures.

A frame sampled two and a half seconds in, where the motion in the scene has picked up. The subject keeps its shape and its colouring across the two ends of the shot, which is the property the publisher describes as identity retention across clips.

A frame sampled four and a half seconds in, at the point where the scene turns into the chaos the published prompt asked for. Read with the two frames above it, it shows one shot opening quiet and ending loud without the subject changing.
Questions
Is this the official HappyHorse 1.1 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.
Why can I not select reference-to-video for this model?
The workspace qualifies a model for reference-to-video by looking for a video-type field on the reference form. This model's reference form takes images only, so the test fails and the tab is disabled. The mode is published and callable on the service side; the limitation is in the interface.
Do the aspect ratios I pick for image-to-video change anything?
No. The request is assembled only from fields the selected mode declares, and image-to-video declares no aspect-ratio field, so the value is dropped. Two independent integrations document the upstream endpoint as rejecting the parameter as well.
How many aspect ratios does this model actually support?
Three answers are in circulation. This catalogue publishes five, the source address it names publishes nine, and the publisher's own worked example shows none. No first-party page publishes a list, so this page reports all three rather than choosing.
What does it cost?
This catalogue publishes no points rule for this product, only an exception fallback of nineteen points per task. Four published price lists exist elsewhere and they do not agree. The server totals the charge for your mode and settings before a task starts, and that total is the one that binds.
Is there a 480P option?
The publisher prices one at CNY 0.45 per second and this catalogue does not publish it. The resolutions here are 720p and 1080p.
Does the model generate audio?
Five sources give five answers and this service gives none, because it exposes no audio control for this model. The publisher's own news entry counts audio among five upgraded dimensions; the source address this catalogue names shows the native audio row as blank. All of them are listed with their origins in the audio section above.
Are the pictures on this page by Alibaba?
No. The publisher publishes no demonstration material on any archivable surface. Every picture here comes from an integration page and is attributed to it, which is why several captions name the platform rather than the model.
Why are two of the images labelled as inputs?
Because the platform supplying them labels them that way. They are the reference images the demonstration clip was made from, so presenting them as output would misstate what they are.
Why is the clip filed under a different name?
The platform serving it places the file under an asset path reading happy-horse-1, the previous generation, while the page hosting it is the current model's playground. The caption says so rather than smoothing it over.
Who is 快乐小马?
It is the publisher's own Chinese name for the same product. Its Chinese consumer site carries it in the page title.
What is missing from this page compared with the other model pages here?
A prompt gallery, because a gallery item needs a prompt for each picture and no source in this field publishes per-image prompts for this model. The page also carries fewer examples than the others, for the same reason: the publisher publishes no demonstration material and only a small part of the third-party pool is attributable.
Explore next
AI video generator
Write a shot, animate a still, or drive a clip with a reference video, then read the quoted credit cost before the job starts.
From prompt to finished image
A worked walkthrough of one image task: choosing a model, writing the prompt, reading the credit quote and judging the result.