A Mystery Model Arrives at the Top
In early April 2026, an AI video generation model with no public track record appeared on the Artificial Analysis Video Arena — the most widely cited blind-comparison benchmark for AI video generators — and immediately claimed the top position.
HappyHorse-1.0 was submitted pseudonymously to the arena, where users compare pairs of generated videos without knowing which model produced them. Within days, the model accumulated enough votes to register Elo scores that surpassed every other entrant, including ByteDance's Seedance 2.0, which had held the top spot since its launch in late March.
The model's arrival drew attention not just for its performance, but for the unusual circumstances around its release. Multiple domains appeared (happyhorse.mobi, happyhorse-ai.com, happy-horse.ai), a press release was distributed through wire services on April 9, and the model's creators initially declined to identify themselves.
Leaderboard Results Across Categories
The Artificial Analysis Video Arena uses an Elo rating system derived from blind user votes. Users compare two videos generated from the same prompt without knowing which model created each one. Higher Elo scores indicate a model is preferred more often. As of April 9, 2026, HappyHorse-1.0 holds these positions:
| Category | Rank | Elo Score | Next Closest |
|---|---|---|---|
| Text-to-Video (no audio) | #1 | 1,333–1,357 | Seedance 2.0 (1,273) |
| Image-to-Video (no audio) | #1 | 1,391–1,406 | Seedance 2.0 (1,355) |
| Text-to-Video (with audio) | #2 | 1,215 | Seedance 2.0 (1,220) |
| Image-to-Video (with audio) | #1 | 1,160 | Seedance 2.0 (1,158) |
For context, the previous consensus top performers — Runway Gen-4.5, Kling 3.0, and Google Veo 3.1 — all score in the 1,100–1,250 range in the text-to-video category. A 60-point Elo gap is substantial in a field where leading models have clustered within 30–40 points of each other.
It is worth noting that Elo scores on community arenas can shift as more votes accumulate. Early ratings based on smaller vote counts tend to be more volatile, and the model's final standing may settle differently over time.
The Alibaba Connection
According to reporting by The Information, HappyHorse-1.0 was anonymously released by Alibaba Group. Separately, press materials distributed on April 9 attribute the model to an "independent AI research team formerly from Alibaba's Taotian Group Future Life Laboratory (ATH-AI Innovation Division)."
The team is reportedly led by Zhang Di, former Vice President of Kuaishou and technical lead of Kling AI — the video generation tool that became one of the most widely used in the space during 2025. If confirmed, this would mean the architect of one of the field's leading models has built what appears to be a stronger successor under a different organizational umbrella.
Whether HappyHorse-1.0 is an official Alibaba product, a semi-independent spin-out, or something else remains unclear. Neither Alibaba nor the Taotian Group had issued a public statement as of April 9. Some community members have speculated the model may be a next-generation version of Alibaba's open-source Wan video model family, tested under a pseudonym before official release, but no direct evidence supports this theory.
Technical Architecture and Capabilities
According to available press materials, HappyHorse-1.0 is built on a 15-billion-parameter unified single-stream Transformer architecture. The key technical claim is that it generates synchronized audio and video in a single forward pass, rather than producing video and audio in separate stages.
Reported capabilities include:
- Native support for Mandarin, Cantonese, and six additional languages with lip synchronization
- Low word error rate for generated speech
- Text-to-video and image-to-video generation
- Inference on a single NVIDIA H100 GPU, with community-optimized versions for consumer GPUs described as "under development"
None of these claims have been independently verified through a published technical paper or third-party evaluation. The leaderboard scores validate output quality as judged by users, but they do not confirm the underlying architecture or specific capability claims.
Open Source Claims vs. Reality
HappyHorse-1.0 has been marketed as "fully open source with complete commercial licensing." Press materials state that "all model weights, distilled models, super-resolution modules, and inference code" are publicly available on GitHub.
As of April 9, 2026, this does not appear to be the case. Multiple independent checks have found:
- GitHub links on official HappyHorse sites return 404 errors or "coming soon" pages
- No model weights appear on HuggingFace
- No inference code has been located in public repositories
This gap between claimed availability and actual accessibility is the most notable concern around the project. It is possible that the release is staged and weights will appear in the coming days, but the current state contradicts the press materials.
For practitioners evaluating AI video models, the distinction matters. Elo scores from blind comparisons are a useful quality signal, but they do not address API availability, pricing, latency, terms of service, or long-term support — all factors that determine whether a model is viable for production use. Until HappyHorse-1.0's code and weights are actually downloadable, its practical value to the broader community remains hypothetical.
The AI video generation space continues to move quickly. With Sora shutting down later this month, Seedance 2.0 gaining traction, and Google aggressively cutting Veo pricing, HappyHorse-1.0 adds another variable to an already competitive field — assuming its creators follow through on their open-source commitments.
Stop scrubbing. Start creating.
Wideframe gives your team an AI agent that searches, organizes, and assembles Premiere Pro sequences from your footage. 7-day free trial.