Anonymous · Arena · 04 · 07 · 2026
No press release. No team attribution. Within days, it took #2 in both Text-to-Video and Image-to-Video on the Artificial Analysis Video Arena. Three days later, Alibaba ATH claimed it. Meet HappyHorse-1.0 — a 15B-parameter unified video + audio model.
Origin story
For three days the AI community tried to guess who built it. The answer was unconventional enough to be its own product strategy.
2026 · 04 · 07
A pseudonymous entry quietly lands on the Artificial Analysis Video Arena leaderboard. No announcement, no team, no paper. Within 48 hours it is trading punches with Veo-3.1 and Grok-Imagine at the top of both video tracks.
blind arena · no attribution
2026 · 04 · 10
Three days after the mystery landing, Alibaba confirms HappyHorse-1.0 as a project from their ATH AI Innovation Unit — a 15B-parameter unified multimodal Transformer that generates cinematic video with natively synchronised audio in a single inference pass.
ath ai innovation unit · confirmed
Arena ELO · April 2026
Blind-tested pairwise on the Artificial Analysis Video Arena. Numbers below reflect public Elo as of mid-April 2026.
#2 — leading #3 Veo-3.1 (with audio) by +69 points.
#2 — leading #3 Grok-Imagine-Video-720p by +23 points.
Gap to the closest head-to-head contender on TTV.
Four technical breakthroughs
Audio. Consistency. Motion. Instruction-following. HappyHorse moves each of these meaningfully — watch each demo and the pattern shows up in a few seconds.
01 / Unified architecture
Most AI video pipelines generate video first and then bolt audio on — a two-stage process that leaks latency and lip-sync drift. HappyHorse-1.0 uses a unified multimodal Transformer that emits synchronised video + audio in one inference pass.
02 / Consistency
Characters, lighting, styles and environments hold across every cut — the highest cross-clip consistency reported by any video model in 2026. That stability is what lets a 20-second scene feel like a film rather than a slideshow of related frames.
03 / Motion & physics
Fewer floaty gestures. Cleaner scene transitions. Physics that holds under scrutiny — objects with weight, fabrics that fold, characters that don't morph between frames. The kind of baseline that professional workflows actually need.
04 / Prompt following
Complex camera moves, specific lighting conditions, nuanced character interactions — HappyHorse-1.0 lands multi-part prompts with notably less re-rolling. Fewer regenerations mean less waste, faster iteration, closer-to-intent outputs.
Technical specifications
Everything Alibaba ATH has disclosed about HappyHorse-1.0 as of April 2026. Weights, pricing and API surface are still pending the official release.
Model size
15 billion parameters
Architecture
40-layer self-attention Transformer · unified multimodal · no cross-attention
Developer
Alibaba — ATH AI Innovation Unit
Output resolution
Up to 1080p
Audio
Native, natively synchronised with video
Multilingual lip-sync
7 languages
Input modes
Text-to-video · Image-to-video · Video-to-video · Image + Text-to-video
Arena position
#2 Text-to-Video & Image-to-Video (Artificial Analysis, April 2026)
Head-to-head
Artificial Analysis maintains four leaderboards — text/image to video, each with or without audio. Noon, 13 April 2026: HappyHorse leads the no-audio tracks with a historic image-to-video high.
Text-to-Video · No Audio
Apr 13 · 2026HappyHorse-1.0
Alibaba-ATH
Dreamina Seedance 2.0 720p
ByteDance Seed
SkyReels V4
Skywork AI
Kling 3.0 1080p (Pro)
KlingAI
grok-imagine-video
xAI
Runway Gen-4.5
Runway
Vidu Q3 Pro
Vidu
Image-to-Video · No Audio
Apr 13 · 2026HappyHorse-1.0
1413
Highest ELO ever recorded on the platform — +56 over Seedance 2.0, on 14,587 samples.
Dreamina Seedance 2.0 720p
ByteDance Seed
grok-imagine-video
xAI
PixVerse V6
PixVerse
Kling 3.0 Omni 1080p (Pro)
KlingAI
SkyReels V4
Skywork AI
Kling 2.5 Turbo 1080p
KlingAI
Veo 3.1 Fast
HappyHorse advantages
What the silent launch signals
Shipping an unattributed model to the top of a blind arena — and waiting three days to claim it — is a confident product strategy. It says: the technology is ready to be judged without the logo. Whether Alibaba repeats that playbook or not, the bar for every other video lab just moved.
Where it earns its keep
Creators & marketers
Cinematic clips at scale without a crew. Social-ready output, native audio, brand-consistent style.
Film & entertainment
Pre-vis, storyboards, final-tier B-roll. 87% cross-clip stability survives a timeline cut.
E-commerce
Product videos with multilingual voiceovers from one prompt — ship localised creative globally.
Education & training
Learning content with accurate lip-sync across seven languages — accessibility by default.
Availability
No public API, no downloadable weights, no confirmed pricing. HappyHorse-1.0 is on top of the arena while its launch surface catches up.
Coming to Yihook at launch
Unified video + audio. 87% cross-clip consistency. Seven-language lip-sync. Available inside Yihook the moment the API goes live — waitlist members get access first.