HappyHorse 1.0
What is HappyHorse 1.0?
HappyHorse 1.0 is Alibaba's video generation model, built by the company's Token Hub (ATH) team and released in April 2026. It renders video and matching audio in a single pass, so dialogue, ambient sound, and on-screen action come back already in sync. One endpoint covers three input modes: text-to-video from a prompt, image-to-video from a first-frame image, and reference-to-video from 1 to 9 character images tagged in the prompt as character1, character2, and so on. Output runs at 720P or 1080P across five aspect ratios (16:9, 9:16, 1:1, 4:3, 3:4), with video lengths from 3 to 15 seconds.
The model first appeared anonymously on the Artificial Analysis video arena in early April 2026 and took the top spot in both the text-to-video and image-to-video blind rankings before Alibaba confirmed it was theirs.
Key features of HappyHorse 1.0
Five features cover how HappyHorse 1.0 fits into a production pipeline.
Joint audio and video in one pass
The picture and the soundtrack render together, so dialogue, footsteps, and ambient noise stay locked to on-screen events. No second TTS or foley pass, no manual lip sync, no drift between two pipelines.
Three input modes on one endpoint
Text-to-video for fresh briefs. Image-to-video for animating a still. Reference-to-video for character continuity. Switch by setting the `mode` field; the rest of the request stays the same.
Up to nine reference images for character work
Reference-to-video takes 1 to 9 images, up to 20 MB each. Tag each one in the prompt as `character1`, `character2`, and so on, and the model holds identity, costume, and silhouette across the video.
Five aspect ratios at 720P or 1080P
16:9, 9:16, 1:1, 4:3, and 3:4 cover horizontal hero, vertical social, square feed, and the older broadcast ratios in one model. Choose 720P for fast drafts or 1080P for delivery.
3 to 15 second videos with seed control
Video length is one integer field with a 5-second default. Pair it with a seed (0 to 2,147,483,647) to get the same render back on re-runs, which makes A/B prompt tests repeatable.
Best for
Ads and marketing with native audio
Cinematic videos from text or product images that come back with matching audio in the same render. No second pass for voice or foley.
E-commerce hero shots
Animate a product still with image-to-video. The first frame anchors the look while the prompt drives the camera move and ambient sound.
Short-form drama and serialized videos
Reference-to-video keeps a cast steady across scenes. Use `character1`, `character2` tags in the prompt to lock identity over a sequence.
Social campaigns with multiple ratios
Five aspect ratios in one model means a 16:9 cut, a 9:16 cut, and a 1:1 cut all come from the same prompt without re-prompting per platform.
Brand-aligned reference work
Up to nine references per call carry brand colours, costume, and product geometry across an entire campaign without training a custom model.
Quick turnaround dialogue scenes
Joint audio means a video with spoken lines or beat-driven cuts works in one call, with no separate sync step at the end.
Use cases
Build a product hero in one call by passing the packaging shot as first_frame_url and a 5-second camera-move prompt; the result lands at 1080P with matching ambient audio. Make a vertical TikTok cut by switching ratio to 9:16 and re-running the same prompt. Run a serialized story by uploading 2 or 3 character references and writing prompts that mention character1 and character2 directly, so the cast stays steady across scenes. Lock a winning result by re-using its seed, so you can re-render at a different duration without losing the look.
Limitations
Aspect ratio control only applies to text-to-video and reference-to-video. Image-to-video has no ratio parameter; output follows the source frame, so start from a still at the target ratio.
Prompts cap at 5,000 non-Chinese characters, or 2,500 Chinese characters. Anything past the cap is truncated, not rejected, so long briefs lose their endings quietly.
There is no audio input parameter. The model writes its own soundtrack; you cannot supply a voiceover file or a music bed.
Output tops out at 1080P, and access is API-only. The open-weights release Alibaba announced at launch has not landed.
HappyHorse 1.0 vs HappyHorse 1.1
HappyHorse 1.1, released June 23, 2026, keeps the same API shape: three modes, 720P or 1080P, 3 to 15 seconds, up to nine reference images, seed, and watermark toggle. Two things changed at the parameter level. Aspect ratio coverage grew from five options to nine, adding 4:5, 5:4, 21:9, and 9:21. And reference-to-video prompts tag images as [Image 1] and [Image 2] instead of 1.0's character1 syntax.
Output quality also moved: 1.1 has smoother motion, steadier subjects, and tighter audio sync. Stay on 1.0 if your prompt library is built around the characterN format; otherwise 1.1 is the better default.
When to use HappyHorse 1.0
Use HappyHorse 1.0 when a video needs picture and sound in one call and the five standard ratios cover your formats. It fits pipelines that already speak the characterN reference syntax, and teams that want one endpoint for prompt-only drafts, animated stills, and multi-reference character scenes. If delivery calls for 21:9 or 9:21, or you want the newest output quality, step up to HappyHorse 1.1.
API examples
Call HappyHorse 1.0 from any language by POSTing to /v1/tasks. Full parameter docs live at docs.unifically.com/models/video/alibaba/happyhorse-1.0-video.
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "alibaba/happyhorse-1.0-video",
"input": {
"mode": "t2v",
"prompt": "A golden retriever running through a field of wildflowers at sunset",
"resolution": "1080P",
"ratio": "16:9",
"duration": 5
}
}'
Successful submission returns a task_id. Poll GET /v1/tasks/<task_id> or set a callback_url on the request to receive the finished video URL.
FAQs
People also ask
HappyHorse 1.0 is Alibaba's video generation model that produces video and matching audio in a single pass. It runs in three input modes (text-to-video, image-to-video, and reference-to-video) on the same endpoint, with output at 720P or 1080P, five aspect ratios, and videos from 3 to 15 seconds.
Three. Text-to-video for prompt-only generation. Image-to-video, which takes a single first-frame image. Reference-to-video, which takes 1 to 9 character images and uses characterN tags in the prompt to anchor each one in the video.
Pass one to nine images on reference_image_urls and refer to each by character1, character2, and so on inside the prompt. The model uses those references to lock character identity, costume, and silhouette across the generated video. Each image can be up to 20 MB.
3 to 15 seconds per video, with 5 seconds as the default. Output runs at 720P or 1080P. Aspect ratios cover 16:9, 9:16, 1:1, 4:3, and 3:4 in text-to-video and reference-to-video modes; image-to-video follows the source frame.
Yes. Audio and video come back in the same render, so on-screen action, dialogue, and ambient sound stay in sync without a second audio pass or manual lip sync afterwards.
Up to 5,000 characters of non-Chinese text or 2,500 Chinese characters per call. Anything past the cap is truncated, so trim long briefs before sending.
Yes. Pass an integer seed in the range 0 to 2,147,483,647 alongside the prompt and any input frames. Re-sending the same inputs and seed returns the same render.
