HappyHorse 1.1: the video model that holds up to nine references
HappyHorse is Alibaba’s video generation model. Its mode list is nearly the same as its neighbour Wan’s, and they differ in details that decide everything: HappyHorse takes the longest outside footage for editing — up to a minute, and it holds up to nine reference images in one clip.
Below is the model itself. Describe a clip or upload a frame: the answer arrives on this page, and the first attempt needs no sign-up.
What you are trying here
This is not a demo mode: the request goes to the same model the main chat uses. A job takes one to five minutes, which is normal; no need to close the page.
Nine references — the reason it gets switched on
One reference holds a character. Three hold the character, their clothes and the object in their hands. Nine is already a small cast: two or three characters, props and the texture of a background, all of which have to match what you supplied rather than be invented afresh.
For comparison: Wan takes ten plus video, Seedance four, and Veo three. If the job sounds like “here are the characters, here is the product, here is the interior — build a clip from these”, the choice is between two engines, and HappyHorse is the one that costs less at 1080p.
References have to be named in words
The model does not guess who is in which picture. In the prompt they are cited in upload order — [Image 1], [Image 2] — and must be described: “the woman in the red dress from [Image 1] picks up the cup from [Image 3]”. Without descriptions the subjects blend together, and the more references there are the more obvious it gets.
Editing outside footage: the gentlest input in the book
Several engines accept a finished file and change something in it, but the conditions differ, and HappyHorse is the least demanding: MP4 or MOV, from three seconds to a minute, up to 100 MB.
| Engine | What it accepts for editing |
|---|---|
| HappyHorse | Any video of yours: MP4/MOV, 3–60 s, up to 100 MB |
| Wan | Your video up to 15 s; output matches it or extends it to 30 s |
| Seedance | Any video of yours, and the only one that stitches several clips |
| Veo | Only its own 720p Veo clip, no older than two days |
It takes a minute in, but will not give a minute back
The input video can be up to sixty seconds, but the output is still capped at fifteen. That is not a setting and not a bug: a long clip has to be edited in pieces and assembled afterwards.
And an edit is billed on the sum of input and output rather than the output alone — so a minute-long source is expensive in itself, even if you only need a short result.
Four modes
| Mode | Input | When you want it |
|---|---|---|
| Text to video | A description only | No frame but an idea. The ratio is set by hand. |
| Frame to video | A single image | Animate a still. The shape comes from the picture and cannot be set separately. |
| References | From 1 to 9 images | Build a scene from your own characters and objects. This model’s main mode. |
| Video editing | A 3–60 s video and up to 5 images | Style transfer or replacing a detail in finished footage. |
One detail about versions
The chapter title says 1.1, and that is true for three of the four modes: text to video, animating a frame and references all run on 1.1. Video editing is still on 1.0 — Alibaba simply has no 1.1 endpoint for it.
The practical consequence: comparing edit quality against generation quality is not a fair comparison, because they are two different generations of the model. When 1.1 editing ships, this paragraph updates and the page address does not change.
How HappyHorse differs from Wan
| HappyHorse 1.1 | Wan 3.0 | |
|---|---|---|
| References per call | Up to 9 images | Up to 10 images, 5 videos and 5 audio clips |
| Clip length | 3–15 s | 2–30 s, and extends existing clips |
| Editing input | MP4/MOV, 3–60 s, up to 100 MB | Up to 15 s |
| Price | Dearer at 720p, cheaper at 1080p | Cheaper at 720p, with a 480p tier at half that; 1080p is double 720p |
What HappyHorse will not do
- It will not take video as a reference. Images only. For reference video you need Wan.
- It will not sync speech to your recording. For a talking character, use Hailuo.
- It will not continue a clip. Extending a finished clip is Wan; stitching several with transitions is Seedance.
- It will not return a clip longer than fifteen seconds — not by generation and not by editing a minute-long source.
- It will not go above 1080p. Higher is only Veo.
How not to overspend
- Count editing as input plus output. This is the main source of surprise bills on this model: a minute-long source is paid for even when you only need five seconds out.
- Drafts at 720p, five seconds. The composition of the scene, and whether the references got mixed up, is visible there.
- Add references one at a time. Nine pictures at once is a reliable way to get mush and not know which one is causing it.
- Editing is free. Trimming, joining, adding audio, pulling a frame — ordinary file operations, not generation.
Frequently asked questions
What is HappyHorse and whose model is it?
How many reference images does HappyHorse take?
Can I edit finished video?
How does HappyHorse differ from Wan?
Why does editing work worse than generation?
How long will the clip be?
Do I need to select HappyHorse in the chat?
Examples and prices are on the video generation page. What we connected and when is in the 1.1 release post. Other models are in the book’s table of contents; video in general is in the video generation book.
Try it yourself
Everything in this guide runs inside Twelver
One chat for text, images, video, music and voice — no separate services or subscriptions.
Open Twelver chat