Happy Horse 1.0 is Alibaba's chart-topping generator that produces 1080p video and its synchronized audio in a single pass — dialogue, Foley, and ambient sound included, no dubbing step.
A unified single-stream Transformer emits video frames and their matching soundtrack jointly in one forward pass, so lip movement, dialogue, and ambience are aligned by design rather than stitched afterward.
Debuting anonymously, it topped the Artificial Analysis Video Arena in both text-to-video and image-to-video on blind human preference votes, overtaking previous leaders.
Generates emotionally expressive, lip-synced dialogue across seven languages including English, Mandarin, Cantonese, Japanese, Korean, German, and French.
Beyond generation, it edits existing footage — restyle a clip or swap in a subject from a reference image while keeping the original motion.
Happy Horse 1.0 (stylized HappyHorse-1.0, Chinese 快乐小马) is an AI video generation model from Alibaba that made an unusual entrance: it appeared at the top of the Artificial Analysis Video Arena leaderboard in April 2026 without any company attached to it, and within days Alibaba confirmed it was behind the model. Built within the Alibaba Token Hub (ATH) business unit by a team led by Zhang Di — a former Kuaishou VP and technical architect of Kling AI — it quickly became one of the most talked-about releases of the year, claiming the top ranking in both text-to-video and image-to-video categories on blind human-preference voting.
What sets Happy Horse 1.0 apart technically is native joint audio-video generation. Rather than producing a silent clip and dubbing sound in a separate step, its Transformer (reportedly a roughly 15-billion-parameter single-stream design, not officially confirmed) generates the video and its synchronized soundtrack — dialogue, Foley, and ambient sound — together in one forward pass. Combined with native 1080p output, multi-shot sequences, and lip-synced speech across seven languages, this makes it aimed squarely at end-to-end, ready-to-watch video rather than raw visuals that need a post-production pipeline.
The model's defining feature is generating picture and sound together in a single pass rather than as two stages. Because audio is produced alongside the frames, lip movement, spoken dialogue, sound effects, and background ambience stay coherent without a separate dubbing or Foley step.
Reports describe a roughly 15-billion-parameter single-stream self-attention Transformer that handles both modalities in one architecture, rather than bolting an audio module onto a separate video generator. This unified design is credited for the tight audiovisual synchronization.
Happy Horse 1.0 generates in HD 1080p natively rather than upscaling from a lower base resolution. It also supports multi-shot sequences, allowing a single generation to move between camera angles within one clip.
The model produces lip-synced, emotionally expressive dialogue across seven languages — English, Mandarin, Cantonese, Japanese, Korean, German, and French. This makes it well-suited to talking-character and localized content where mouth movement must match the spoken track.
It accepts text prompts, still images, reference images and existing video as input. Text-to-video builds a scene from a description, image-to-video animates an existing frame or character design, reference-to-video carries a subject from reference images into a new scene, and video editing reworks existing footage.
On the Artificial Analysis Video Arena, which ranks models by blind human preference, Happy Horse 1.0 reached the top position in both text-to-video and image-to-video, ahead of previously leading models. It is one of the highest-rated video models on that leaderboard to date.
Happy Horse 1.0 is a closed-source, proprietary model: its weights are not publicly available and it is not offered under an open license. You can use it through Alibaba Cloud Model Studio's API, the official HappyHorse website, the Qwen App, and API partners.
| Name | Happy Horse 1.0 (HappyHorse-1.0) |
|---|---|
| Developer | Alibaba (Alibaba Token Hub, ATH) |
| Parameters | ~15 billion (reported, not officially confirmed) |
| Architecture | Single-stream self-attention Transformer (reported) |
| Access | Proprietary; API via Alibaba Cloud Model Studio |
| Inputs | Text prompt, image, reference images, video (editing) |
|---|---|
| Output resolution | 1080p HD (native) |
| Audio | Native synchronized audio (dialogue, Foley, ambience) |
| Multi-shot | Supported within a single clip |
| Lip-sync languages | English, Mandarin, Cantonese, Japanese, Korean, German, French |
| Leaderboard | Artificial Analysis Video Arena |
|---|---|
| Text-to-video rank | #1 (blind human preference) |
| Image-to-video rank | #1 (blind human preference) |
| Debut | April 2026 (anonymous, later attributed to Alibaba) |
Because lip-sync and speech are generated with the video, the model suits avatars, explainers, and character scenes where mouth movement must match spoken lines — across seven languages without a separate voiceover pass.
Native audio means short-form clips arrive with ambience and sound effects already attached, cutting the editing loop for creators publishing quickly to social platforms.
Multilingual lip-sync makes it practical to produce the same scene in several languages with matching mouth movement, useful for global marketing and education content.
Image-to-video lets designers and studios bring a single frame, product shot, or character illustration to life with motion and a synchronized soundtrack.
Multi-shot 1080p generation with sound helps teams sketch out sequences, mood, and pacing early — a fuller preview than silent draft renders.
Happy Horse 1.0 overtook ByteDance's Seedance 2.0 at the top of the Artificial Analysis Video Arena. Here's how the two leading models line up on the headline capabilities.
| Feature | Seedance 2.0 | Happy Horse 1.0NEW |
|---|---|---|
| Video Arena rank (T2V, at HH debut) | Previous leader | #1 |
| Native synchronized audio | Yes, joint audio-video generation | Yes, joint audio-video in one pass |
| Multilingual lip-sync | Supported | 7 languages |
| License / access | Closed / commercial | Closed / API |
| Developer | ByteDance | Alibaba |
Happy Horse 1.0 is closed source: its weights are not publicly available, so it cannot be downloaded or self-hosted. Claims of an Apache-2.0 release came from websites not affiliated with Alibaba; access is through Alibaba's own channels and API partners.
Reported Elo scores and exact rankings differ between leaderboard captures and categories. Treat specific point values as approximate; the consistent takeaway is a top-tier, #1-class placement rather than one fixed number.
Like most current video models, generations are short clips rather than long-form video, and complex multi-shot prompts can still produce continuity or artifact issues. Plan for iteration on demanding scenes.
First-party API access is available on Alibaba Cloud Model Studio, and a newer HappyHorse 1.1 has since been released, so pricing and features can differ between 1.0, 1.1 and partner platforms.
Generate 720p or 1080p video with synchronized sound on SharkFoto — set any length from 3 to 15 seconds and start from text, an image, or reference images.
Try Happy Horse 1.0 now