When we covered HappyHorse-1.0 on April 13, the model had just claimed #1 on the Artificial Analysis Video Arena — but no public API existed. That changed on April 26, 2026, when fal launched developer and enterprise access to HappyHorse-1.0, four days ahead of the originally expected date.
The model is now accessible to anyone building with AI video generation.
What Happened on April 26
The timing was notable. At 9 PM PST on April 26 — the same day OpenAI shut down the Sora web app — fal went live with HappyHorse-1.0. Whether coincidental or deliberate, the contrast was sharp: one product closes, another opens its API the same evening.
HappyHorse-1.0 is built by Alibaba’s Taotian Future Life Lab and holds an ELO score of 1,381 on the Artificial Analysis blind leaderboard — 107 points above the second-place model. In blind human comparisons, users prefer its output roughly 65% of the time.
Technical Architecture
HappyHorse-1.0 uses a 15-billion-parameter self-attention Transformer with 40 layers. The architecture generates video and audio jointly in a single forward pass — not two separate models piped together. This results in naturally synchronized dialogue, ambient sound, and effects without additional post-processing.
Key specs:
- Resolution: 720p or 1080p
- Clip length: 3 to 15 seconds
- Aspect ratios: 16:9, 9:16, 1:1, 4:3, 3:4
- Languages for lip-sync: English, Mandarin, Cantonese, Japanese, Korean, German, French
- Generation time: ~38 seconds per 1080p clip on a single NVIDIA H100
Four API Endpoints
fal exposes HappyHorse-1.0 through four endpoints:
- text-to-video — generate a clip from a text prompt
- image-to-video — animate a still image
- reference-to-video — generate video while maintaining character or subject consistency
- video-edit — apply edits to an existing clip
The reference-to-video endpoint is particularly useful for multi-shot sequences where character identity must remain consistent — historically one of the harder problems in AI video generation.
Commercial Rights
fal has confirmed that all outputs generated through the HappyHorse-1.0 API carry full commercial rights.
Open Source Confusion
There have been claims circulating that HappyHorse-1.0 is open source. As of the publication date of this article, this is not the case.
No code, model weights, or reproducible implementation have been released publicly. The model is only accessible via API.
Where Tellers Stands
HappyHorse-1.0 is already integrated internally at Tellers.
We can provide access on request, and are currently finalizing our evaluations before including it by default in the Tellers agent.
For AI video creation within a full editing workflow, Tellers already supports Runway Gen 4.5, LTX Video, Kling, Hailuo, Veo 3.1, and more. You can generate clips, edit on a timeline, combine with stock footage, and export — all through one workflow or via the Tellers API.
What the ELO Gap Actually Means
A 107-point ELO margin is significant. Because ELO operates on a logarithmic scale, this gap translates to roughly a 65% preference rate in blind comparisons.
However, leaderboard performance does not guarantee superiority across all use cases. Different prompt types — especially structured edits, UI-heavy scenes, or long-form coherence — may yield very different results.
That is why real-world evaluation matters more than rankings alone.
FAQ
Is HappyHorse-1.0 available on Tellers?
Yes — on request. We have already integrated HappyHorse-1.0 internally and can provide access to users upon request. We are finalizing our evaluations before rolling it out in the default Tellers agent.
Is HappyHorse-1.0 open source?
No — despite some claims online, there is currently no public release of code or model weights. As of the date of this article, nothing has been made available.
What video models does Tellers currently support?
Tellers supports Runway Gen 4.5, LTX Video (with first and last frame control), Kling, Hailuo, Veo 3.1, and more. The full list is visible in the app.
What resolution does HappyHorse-1.0 generate?
HappyHorse-1.0 outputs at 720p or 1080p for clips from 3 to 15 seconds, across multiple aspect ratios including 16:9, 9:16, and 1:1.
Does HappyHorse generate audio alongside video?
Yes. Audio is generated in the same forward pass as the video — there is no separate audio model or post-processing step.
Which languages does HappyHorse-1.0 support for lip-sync?
English, Mandarin, Cantonese, Japanese, Korean, German, and French.
What is HappyHorse-1.0 particularly good at?
In our evaluations so far, it performs especially well on 2D animations — a category where many current models still struggle with motion artifacts. These artifacts can sometimes be partially mitigated in other models by adjusting output frame rates, but HappyHorse appears more stable out of the box.
When was HappyHorse-1.0 revealed to be an Alibaba model?
On April 10, 2026, after appearing anonymously on the Artificial Analysis Video Arena leaderboard on April 7.
The current top-ranked model in AI video generation is now accessible via API.
It shows strong performance — particularly in animation stability — but still requires careful evaluation depending on your use case.
If you want early access, we can enable it on Tellers. And if you need a complete workflow — generation, editing, orchestration across models, and API access — start on Tellers.