HappyHorse AI is the open-source video model family from Alibaba's Taotian Group. HappyHorse 1.0 appeared anonymously on the Artificial Analysis Video Arena in April 2026 and took #1 in both text-to-video and image-to-video within days. HappyHorse 1.1 is the current version, with the same native audio and a lower 1080p price.
On Modellix, HappyHorse 1.1 costs $0.14 per second at 720p and $0.18 at 1080p, with 3 to 15-second clips and native audio. It fits realistic people, product shots, and fast motion where you want audio in the same pass.
If you need a single 30-second take, or the top audio score on the arena, Seedance is still the stronger pick. The comparison below shows where each one wins.
HappyHorse AI at a Glance
| HappyHorse | |
|---|---|
| Position | Alibaba Taotian's open-source video line; 1.1 is the current, streamlined version of 1.0 |
| Built by | Future Life Lab, Alibaba Taotian Group, led by Zhang Di (formerly the technical lead behind Kling) |
| Best for | Realistic people and hands, product and food shots, fast sports motion, wildlife |
| Routes on Modellix | 1.1 T2V, 1.1 I2V, 1.1 R2V (1 to 9 reference images), plus the 1.0 routes and 1.0 Video Edit |
| Price (1.1) | $0.14/s at 720p, $0.18/s at 1080p |
| Clip length | 3 to 15 seconds |
| Audio | Generated in the same pass as the video |
| Aspect ratios (1.1) | 16:9, 9:16, 1:1, 4:3, 3:4, plus 4:5, 5:4, 9:21, and 21:9 |
| Open weights | 1.0 weights, distilled versions, and inference code released on GitHub under a commercial license |
Specs from the Modellix model pages and the HappyHorse team's April 2026 release.
HappyHorse 1.1 vs 1.0
| HappyHorse 1.0 | HappyHorse 1.1 | |
|---|---|---|
| 720p price | $0.14/s | $0.14/s |
| 1080p price | $0.24/s | $0.18/s |
| Clip length | Up to 15 s | 3 to 15 s |
| Aspect ratios | Standard set | Adds 4:5, 5:4, 9:21, and 21:9 |
| Reference-to-video | Up to 9 images | 1 to 9 images |
| Video editing | 1.0 Video Edit route | Use 1.0 Video Edit |
The practical change is price at 1080p: 1.1 is 25% cheaper per second, and the wider aspect ratios cover social feeds and cinematic banners without cropping. Default to 1.1 for new work, and keep 1.0 only for the Video Edit route or a prompt set you have already tuned on it.
HappyHorse Pricing
| Route | 720p | 1080p | 5 s at 1080p | 15 s at 1080p |
|---|---|---|---|---|
| HappyHorse 1.1 (T2V, I2V, R2V) | $0.14/s | $0.18/s | $0.90 | $2.70 |
| HappyHorse 1.0 (T2V, I2V, R2V) | $0.14/s | $0.24/s | $1.20 | $3.60 |
| HappyHorse 1.0 Video Edit | $0.84/s | $1.44/s | $7.20 | $21.60 |
Clip costs are our calculations from the per-second rates. Billing is per generated second with no subscription.
For context at 1080p, Seedance 2.0 T2V is $0.37 per second, Wan 3.0 T2V is $0.20, and Kling V3 T2V with audio is $0.1344.
HappyHorse Benchmarks
Scores from the Artificial Analysis Video Arena at launch (April 2026), where users vote between two unlabeled clips from the same prompt.
| Category | HappyHorse 1.0 Elo | Rank | Previous #1 |
|---|---|---|---|
| Text-to-video, no audio | 1,333 to 1,347 | #1 | Seedance 2.0 (1,273) |
| Image-to-video, no audio | 1,391 to 1,406 | #1 | Seedance 2.0 |
| Text-to-video, with audio | 1,205 | #2 | Seedance 2.0 leads by 14 |
| Image-to-video, with audio | 1,161 | #2 | Seedance 2.0 leads by 1 |
A 60 to 74-point lead means HappyHorse won roughly 58 to 60% of head-to-head votes without audio. With audio included, Seedance 2.0 stays on top.
New entries carry more volatile Elo than models with thousands of votes, so the gap can narrow over time. Check the live leaderboard before treating these as current rankings.
How HappyHorse Generates Video
- One Transformer for every modality. Text, image, video, and audio tokens run through a single 40-layer self-attention stack, so video and sound come out of the same forward pass instead of a separate audio model.
- 15B parameters, 8 denoising steps. Distillation keeps inference short; the team reported about 38 seconds for a 1080p clip on one H100.
- Native lip sync in seven languages. Mandarin, Cantonese, English, Japanese, Korean, German, and French.
These figures come from the team's April release. The open weights let you verify them on your own hardware.
What Can You Build with HappyHorse
HappyHorse is at its best when a shot has to look filmed rather than generated.
- Product and food ads. Close macro shots with liquid, steam, and texture, rendered with sound in one call.
- UGC-style and lifestyle clips. Natural faces and hands for testimonial-style or social content.
- Sports and action b-roll. Fast body motion with realistic motion blur.
- Documentary and nature footage. Animals and landscapes that hold up at 1080p.
Each of those rests on how the model handles faces, fast motion, fine texture, and natural movement. Here is how HappyHorse 1.1 renders each one.
Copy any prompt below and try it yourself in the HappyHorse 1.1 T2V playground. Generated on Modellix: 5 seconds, 720p.
Photoreal human close-up
“Close-up of an elderly fisherman mending a net on a wooden dock at golden hour, weathered hands, a soft smile, gentle sea breeze.”
Fast sports motion
“A skateboarder lands a kickflip down a flight of city stairs, low-angle tracking shot, natural motion blur, afternoon sun.”
Food & product detail
“Honey pours slowly over a stack of pancakes topped with berries, steam rising, macro shot, bright breakfast light.”
Wildlife documentary realism
“A snow leopard walks along a rocky ridge in falling snow, its breath visible in the cold air, telephoto documentary shot.”
Where HappyHorse is the wrong choice
- You need one continuous take longer than 15 seconds. Seedance 2.5 T2V and Wan 3.0 T2V reach 30 seconds.
- Audio quality is the deciding factor. Seedance 2.0 leads the arena's audio categories.
- You need 4K output. HappyHorse stops at 1080p.
Third-party testers also flagged physics slips in the team's own demos, such as a vehicle moving underwater with no bubbles. Run your hardest shot on both HappyHorse and Seedance before you commit.
Who Built HappyHorse
HappyHorse launched with no company name, no paper, and no launch post. The community spent two days testing theories, including links to Kuaishou and to the open-source daVinci-MagiHuman project.
On April 9, 2026, the team confirmed it is the Future Life Lab inside Alibaba's Taotian Group, led by Zhang Di, who previously led the technical build of Kling 1.0 and 2.0 at Kuaishou. The weights, distilled versions, super-resolution module, and inference code were released the same day.
Third-party sites such as happyhorse.app and happyhorse.video use the name but have not confirmed they run this model. Treat them as separate products.
How to Try HappyHorse in 30 Seconds
Sign in to Modellix with email, Google, or GitHub, add credit, and open the HappyHorse 1.1 T2V model page. Paste any prompt from the clips above and compare your result.
When the result is worth repeating, move to code. Paste this into Claude Code, Codex, or Cursor:
My Modellix API key is mdlx-••••••••••••••••••••••••. Read https://docs.modellix.ai/get-started/index.md and https://docs.modellix.ai/alibaba/happyhorse-1-1-t2v.md, then use HappyHorse 1.1 T2V to generate one 5-second 1080p 16:9 clip for each prompt in prompts.txt. Poll each task until it finishes and save the MP4 files to ./clips.
Swap the masked key for your own from the Modellix console. For raw requests, see the HappyHorse 1.1 T2V docs; the Skill and CLI guides cover agent and terminal workflows.
Modellix collection
Shortlist AI Video Models for Production
Compare text, image, and reference-led video workflows across multiple models from one production surface.
Good first tests: Seedance 2.0 Wan 2.7 HappyHorse 1.1
Explore the collectionFrequently Asked Questions About HappyHorse AI
Who makes HappyHorse AI? The Future Life Lab in Alibaba's Taotian Group, led by Zhang Di. The team confirmed its identity on April 9, 2026.
Is HappyHorse open source? Yes. The 1.0 weights, distilled versions, super-resolution module, and inference code are on GitHub under a commercial license. Check the license file before deployment.
How much does HappyHorse cost? On Modellix, HappyHorse 1.1 is $0.14 per second at 720p and $0.18 at 1080p. HappyHorse 1.0 is $0.14 and $0.24.
Should I use HappyHorse 1.1 or 1.0? Use 1.1 for new work: it is cheaper at 1080p and supports more aspect ratios. Keep 1.0 for the Video Edit route.
How does HappyHorse compare with Seedance 2.0? HappyHorse led Seedance 2.0 in the arena's no-audio categories at launch; Seedance 2.0 led with audio. HappyHorse 1.1 is about half the price of Seedance 2.0 at 1080p.
Can I run HappyHorse locally? The team reports one NVIDIA H100 for about 38 seconds per 1080p clip. Consumer GPU versions were described as in development.
Try HappyHorse
Create a Modellix key, run HappyHorse 1.1 on your own prompt, and pay only for the seconds you generate.
Get API KeyRelated articles
HappyHorse 1.0 API: Integration Guide, Pricing & Code
HappyHorse 1.0 is GA on Alibaba Cloud Model Studio and API aggregators today. Working model IDs, request schema, Python and curl samples, and same-day pricing.
AI Product Video Generator [Tested]: 1 Photo, 6 Models
See real AI product video clips, the cost per usable clip, and the prompt that stops a model from reshaping your product, before you spend a cent.
Wan 2.2 API: Alibaba Video Models on a Unified Key
Wan 2.2 API explained with live 2026 pricing: which model ID to call, t2v vs i2v routes, per-second costs, the async task lifecycle, and one key for the family.
Wan 2.7 API Guide: Pricing, T2V and I2V Access (2026)
Wan 2.7 ships text-to-video and image-to-video APIs. See per-second pricing, the t2v vs i2v split, async setup, and how to access it via Modellix.