For three weeks in April 2026, the model labeled "happyhorse-1.0" sat at the top of the Artificial Analysis Video Arena leaderboard with a 107-point Elo margin over second place — and nobody knew who built it. On April 10, Alibaba's ATH-AI Innovation Division confirmed it was theirs ("快乐小马" — Happy Horse). The fal API opened on April 26. We had it live on Oakgen on April 29 — same day. This review covers what HappyHorse 1.0 actually does well after a week of generations, where it falls behind Seedance 2.0, Veo 3, Kling 3.0, and Sora 2, and whether it deserves the top spot.
HappyHorse 1.0 is live on Oakgen's AI Video Generator. 1,000 free credits to start, no credit card required.
The Stealth #1 Story
HappyHorse 1.0 first appeared on the Artificial Analysis leaderboard on April 7, 2026, listed only by codename. Within 48 hours it had pulled clear of Seedance 2.0, Veo 3, Sora 2, and Kling 3.0 in blind side-by-side voting. The arena format is brutal — users see two clips generated from the same prompt and pick the better one — so a 107-point Elo gap is not a marginal lead. It is the largest single-model jump on the leaderboard since Sora launched.
Alibaba waited three days before claiming the model on X, then another sixteen days before opening the fal API. The delay let the leaderboard validate the model on its merits before any branded marketing push. For a Chinese lab competing with OpenAI, Google DeepMind, and ByteDance for Western mindshare, that was a smart play.
Final aggregate ranking when the API opened: #1, 1381 Elo, 107-point margin over the next model. That is the headline.
Architecture: One Forward Pass for Video and Audio
What makes HappyHorse 1.0 technically interesting is that it does not look like its competitors under the hood.
Most current video models — Seedance 2.0, Veo 3, even the open-source Wan family — generate video and audio as two coupled but distinct processes. There is a video transformer producing frames, and a paired audio model (sometimes a separate diffusion model, sometimes a vocoder) producing the soundtrack. The two communicate through cross-attention layers or shared latent representations, but they are architecturally separate components.
HappyHorse 1.0 is a single-stream 40-layer Transformer with roughly 15 billion parameters. Video tokens and audio tokens flow through the same attention stack. There is no cross-attention between modalities because there are no separate modalities — the model treats a clip as one tokenized sequence that happens to include both visual and acoustic information.
Alibaba's published technical notes claim this gives them three things:
- Tighter audio-visual coupling — every audio token can attend directly to every video token in every layer, not just at designated cross-attention points.
- Faster inference — no separate model passes, no late-stage compositing. Roughly 38 seconds to produce a 1080p clip on a single H100.
- Better multilingual lip-sync — because phoneme timing and mouth shapes are co-generated rather than retrofitted, the model handles seven languages natively (English, Mandarin, Cantonese, Japanese, Korean, German, French).
Whether the single-stream design is responsible for the leaderboard performance, or whether the leaderboard performance just reflects 15B parameters and good training data, is not something we can prove from outside. But the speed claim holds up in our testing — a 5-second 1080p clip averages around 10 seconds end-to-end on Oakgen, against 14–17 seconds for comparable Seedance 2.0 generations.
What HappyHorse 1.0 Is Genuinely Good At
After a week of running prompts through the model on Oakgen, four strengths stand out.
1. Native audio that actually sounds produced
Most models with "native audio" generate something between a sound effect and a placeholder — footsteps that sound like dry leaves, ambient music that loops too obviously, dialogue that betrays the AI origin within a syllable. HappyHorse's audio is the first we have heard from a video-first model that you could ship without a sound designer touching it.
A test prompt for a rainy night street scene produced rain hitting different surfaces at different volumes, distant traffic with occasional horns, and footsteps that matched the character's gait and the shoe-on-pavement physics. None of that was prompted explicitly — it came from the visual context.
2. Multilingual lip-sync across 7 languages
This is where HappyHorse pulls clear of most Western models. Generating the same 8-second monologue prompt in English, Mandarin, Japanese, and German returned four clips where the mouth shapes matched the phonemes of each language. Veo 3 still has tighter sub-10ms English dialogue lip-sync, but its non-English performance drops off fast. HappyHorse holds quality across all seven supported languages.
For agencies working on global ad campaigns, that is a real workflow change. Generate once per language instead of generating in English and dubbing.
3. Image-to-video without character drift
Upload a portrait or product shot, write a prompt, and HappyHorse maintains identity across the full 15 seconds. Faces stay coherent, fabric textures hold, branded packaging keeps its label. This is the category where it scores 1401 Elo on the Artificial Analysis arena — the highest single-category score on the board.
4. Speed that matters for iteration
Roughly 30–40% faster than Seedance 2.0 at the same resolution. On a per-clip basis that sounds modest. Across a session where you generate 20 variations of a hero shot, it adds up to ten or fifteen minutes saved.
Real Prompts We Tested
The model accepts text-only and image+text. Here is a prompt that produced one of our best test outputs:
A woman in a red wool coat walks down a snow-covered Tokyo side street
at dusk, neon signs glowing through light snowfall. She stops, turns
to camera, and says in Japanese: "また明日." Camera tracks at waist
height, 35mm lens, shallow depth of field, ambient soundscape with
distant traffic and snow muffling the city.
Length: 8 seconds. Resolution: 1080p.
What HappyHorse delivered: tracking shot held smoothly, snow accumulation on the coat read correctly across the duration, the spoken Japanese lip-sync was clean, and the audio mix had the muffled-by-snow quality the prompt asked for. Total generation time: 11 seconds.
A second test for image-to-video, anchored on a product photo of a ceramic mug:
[image: ceramic mug, white background, three-quarter angle]
The mug rotates slowly clockwise on a dark walnut surface. Steam
rises from the rim and curls in the warm side-lit air. The camera
pushes in over 6 seconds to a tight macro on the steam catching the
light. Audio: gentle ambient room tone, faint creak of the wood as
the mug settles.
The mug's identity held through the rotation. The steam physics looked plausible — not over-animated, not under-animated. The "wood creak" was a small flourish but the model included it.
These are the kinds of outputs you get on the first or second try, not the tenth. That is the bigger story than the leaderboard score.
Generate HappyHorse 1.0 Videos Now
No region restrictions, no business email needed. Start with 1,000 free credits.
Specs at a Glance
| Feature | Specification | HappyHorse 1.0 |
|---|---|---|
| Maker | Alibaba ATH-AI Innovation Division | |
| Architecture | Single-stream 40-layer Transformer, ~15B parameters | |
| Output Resolution | Native 1080p HD | |
| Max Length (Lite tier) | 12 seconds | |
| Max Length (Paid tier) | 15 seconds | |
| Generation Speed | ~10 seconds avg per clip; ~38s for 1080p on a single H100 | |
| Speed vs Seedance 2.0 | ~30–40% faster | |
| Native Audio | Yes — generated in the same forward pass as video | |
| Lip-Sync Languages | 7 — English, Mandarin, Cantonese, Japanese, Korean, German, French | |
| Input Modalities | Text, image | |
| Output Modalities | Video + native audio | |
| Aggregate Arena Ranking | #1 — 1381 Elo, 107-point margin over #2 |
Where HappyHorse 1.0 Falls Behind
This is the part most reviews skip. HappyHorse is the top-ranked model on the leaderboard, but "top of the aggregate ranking" is not "best at every category." Four real weaknesses to know about before you commit to it for a production workflow.
Seedance 2.0 wins Image-to-Video with audio
On the arena's "Image-to-Video with audio" category, Seedance 2.0 leads HappyHorse 1.0 1182 to 1167. Narrow, but consistent across many votes. If your workflow is "drop in a product photo, get a clip with synchronized audio," Seedance is still the model to beat. Most likely because Seedance's @ reference system gives users finer control over how the source image translates to motion, and that control matters more than raw model capability for image-anchored work.
Kling 3.0 generates 4K, HappyHorse caps at 1080p
HappyHorse is locked at native 1080p. For social, web, and most ad delivery, 1080p is fine — that is what the platforms ingest anyway. But for billboards, festival submissions, large-screen pitches, or anywhere a client demands true 4K masters, you have to upscale. Kling 3.0 generates 4K natively. If your output destination is genuinely 4K, Kling is the better starting point even though it loses to HappyHorse in the blind arena.
Sora 2 supports longer clips
Sora 2 generates up to 20 seconds in a single pass. HappyHorse caps at 15 seconds on the paid tier, 12 on the Lite tier. For storytelling shots that need a longer beat — a slow reveal, a sustained emotional moment, a continuous walking-and-talking scene — Sora's extra duration is meaningful. You can chain HappyHorse clips, but chaining introduces continuity errors that a single 20-second generation avoids.
Veo 3 still has tighter dialogue lip-sync for spoken English
If your specific workflow is "AI-generated talking-head explainer in English with synchronized speech," Veo 3 remains marginally better. Veo 3 hits sub-10ms phoneme-to-mouth-shape latency on English dialogue, and the difference is visible in extreme close-ups. HappyHorse's lip-sync is excellent across seven languages, but Veo 3's English performance is the current ceiling. The tradeoff: Veo 3 is much weaker than HappyHorse on non-English languages, and it does not match HappyHorse on environmental audio quality.
Documentation is thin
Worth flagging because it affects day-to-day usage. The model launched on the fal API on April 26 and Alibaba has not yet published an extensive prompt library, parameter reference, or known-failure-mode catalog. If you are coming from Veo or Sora's mature documentation ecosystems, expect to do more empirical testing to find what HappyHorse responds to. The community will catch up — for now, plan for some discovery work.
Who Should Use HappyHorse 1.0
After a week of testing, here is our honest read on fit.
Use HappyHorse 1.0 if:
- You need native audio that sounds produced rather than placeholder, and you do not want to layer on a separate ElevenLabs or Suno pass for ambient sound.
- You are working on multilingual content — global ad campaigns, multi-region product launches, content for non-English markets. Seven-language lip-sync in a single pass is the killer feature.
- You want fast iteration cycles. Roughly 10-second generations let you run 20–40 variations in a session without losing a morning.
- You are doing image-to-video without audio. Top arena score in that category (1401 Elo).
Stick with another model if:
- Your output destination is true 4K. Use Kling 3.0.
- You need clips longer than 15 seconds in a single take. Use Sora 2.
- Your specific deliverable is English-only talking-head with extreme close-up lip-sync. Use Veo 3.
- Your workflow depends on video reference inputs for camera or motion replication. HappyHorse accepts text and image only — Seedance 2.0's @ reference system handles video references natively.
The good news on Oakgen is you do not have to commit. The same 1,000 free credits work across HappyHorse, Seedance 2.0, Kling 3.0, Veo 3.1, Wan 2.6, and 25+ other video models. Run the same prompt through three or four of them, see which output you prefer, and only spend on the variations of the winner.
Earn 25% recurring on every referral.
Share Oakgen, get paid every month they stay.
Verdict
HappyHorse 1.0 earned its #1 spot on the leaderboard. The single-stream architecture appears to be paying off in tighter audio-visual integration, the multilingual lip-sync is genuinely ahead of Western competitors, and the speed is high enough to change how you iterate. It is not a clean sweep — Seedance leads on image-to-video with audio, Kling has higher resolution, Sora goes longer, Veo has tighter English dialogue. But across the broader ranking that combines text-to-video, image-to-video, audio quality, and overall coherence, HappyHorse is the model to beat right now.
For most creators in 2026, the right move is not to pick one model. It is to keep access to all of them and route each shot to whichever wins for that specific need. That is what Oakgen is built for, and it is why HappyHorse 1.0 was live within three days of the fal API launch — alongside the 30+ other video models you already have access to under the same credit balance.
What to Read Next
- HappyHorse 1.0 vs Seedance 2.0: Which AI Video Model Wins in 2026? — Deep head-to-head with side-by-side category-by-category breakdowns.
- HappyHorse 1.0 Prompting Guide: How to Get Cinematic Results in 2026 — Prompt templates, structure, and what actually steers the model.
- How to Make a Game Trailer with HappyHorse 1.0 + GPT-Image-2 — Full pipeline walkthrough: keyframes, motion passes, soundtrack.
