Happy
Horse
The AI video model that appeared from nowhere, topped every leaderboard, and vanished — leaving only questions.
WHAT MAKES IT DIFFERENT
Architecture that breaks the rules
40-Layer Single-Stream Transformer
No cross-attention. Text, video, and audio tokens in one unified sequence. 4 modality layers at each end, 32 shared.
Joint Video + Audio Generation
Dialogue, ambient sound, and Foley alongside video frames in a single pass. No post-production dubbing pipeline.
6 Languages · Lip-Sync
Chinese, English, Japanese, Korean, German, French. Ultra-low WER multilingual lip-sync for digital human content.
LEADERBOARD
Blind-tested. Elo-rated. No self-reporting.
Source: Artificial Analysis Video Arena — live leaderboard data
DEMO
See it in action
The official HappyHorse demo at happyhorse.app — generate videos with text prompts. No sign-up required.
THE MYSTERY
Who built HappyHorse?
Nobody has claimed credit. Artificial Analysis called it “pseudonymous.” It appeared, dominated, then vanished from the leaderboard days later.
Alibaba Wan 2.7?
Same timing. "Horse" = Year of the Horse. But Wan 2.7 focuses on image/reasoning — architecture doesn't match.
ByteDance Seedance?
Top Chinese contender. But Seedance leads in audio. No motive to submit a stronger model under a fake name.
New Unknown Team
Open-source bundle = research playbook. "Rank first, release later" strategy validated by Asian labs. Most likely.
ANALYSIS
Latest Research & Guides
AI Video Model License Guide: What You Can Actually Do Commercially
Open source doesn't always mean free to use commercially. The licensing landscape for AI video models is more nuanced than most developers realize. Here's what matters for production use.
2 min readBest Open Source AI Video Generation Models Compared (April 2026)
The AI video generation landscape shifted dramatically in early 2026. Models that were state-of-the-art in late 2025 are now mid-tier at best. Here's where things stand based on Artificial Analysis Vi...
2 min readHappyHorse 1.0: The AI Video Model That Came From Nowhere
In early April 2026, a model called HappyHorse-1.0 appeared on the Artificial Analysis Video Arena leaderboard under a pseudonymous identity. Within hours it sat at #1 in both text-to-video and image-...
2 min read
The weights haven't dropped yet.
When they do, you'll know.
We track the GitHub, HuggingFace, Arena, and community signals. Bookmark this page.
Read the Full Analysis →