

HappyHorse-1.0
#33 in Open-Source-Sprachmodellealibaba · v1.0 · seit April 2026 · 19× · zuletzt 30. Juni 2026
HappyHorse-1.0 ist ein Videogenerierungsmodell von Alibaba (ATH Innovation Unit), das im April 2026 anonym auf der Artificial-Analysis-Plattform debütierte und dort Rang #1 bei Text-to-Video- und Image-to-Video-Benchmarks erreichte. Das Modell basiert auf einer einheitlichen 40-schichtigen Transformer-Architektur mit 15 Milliarden Parametern und generiert Video und Audio in einem einzigen Forward Pass, ohne separate Audio-Nachbearbeitung. Es wurde am 27. April 2026 offiziell gestartet und unterstützt native Mehrsprachenlip-Sync sowie verschiedene Generierungsmodalitäten (Text-to-Video, Image-to-Video, Subject-to-Video).
Features
| Key-Benchmark (%) | Artificial Analysis Video Arena: Elo 1.333 T2V (no audio, ~+60 Elo vs. Seedance 2.0); Elo 1.392–1.416 I2V (no audio) — Platz 1 in beiden Kategorien (April 2026) |
| Lizenz | Proprietär (geschlossenes API-Modell); keine öffentlichen Weights; open-source-Release angekündigt, aber noch nicht erfolgt (Stand April 2026) |
| Multimodalität | Text-to-Video, Image-to-Video, Reference-to-Video (bis 5–9 Bilder), Video-Editing; natives Audio (Dialogue, Ambient, Foley) im selben Forward-Pass; 7-sprachiges Lip-Sync (EN, ZH, YUE, JA, KO, DE, FR); Output: 720P/1080P, 3–15s, Seitenverhältnisse 16:9/9:16/1:1/4:3/3:4 |
| Plattform | happyhorse.com (Endnutzer), Alibaba Cloud Model Studio / Bailian (API), Qwen App (Consumer), fal.ai (API-Partner) |
| Preis pro 1M Token | Preismodell: pro Sekunde Video-Output. Alibaba Cloud Model Studio: $0.14/s (720P) bis $0.24/s (1080P); fal.ai: $0.14/s (720P) / $0.28/s (1080P) |
| Release-Datum | 27. April 2026 (limitierte Beta / Graustufen-Test-Start) |