Happy Horse 1.0: The Mystery AI Video Model That Shocked Everyone

An unknown model named Happy Horse 1.0 just claimed #1 on the world's biggest blind AI video benchmark — and its creator has finally stepped forward.

On April 7, 2026, a previously unknown AI video generation model called Happy Horse 1.0 appeared at the top of Artificial Analysis's Video Arena, simultaneously claiming the #1 spot in both Text-to-Video and Image-to-Video categories. No team had claimed ownership. No weights were publicly available. No API existed. For three days, the AI community scrambled to figure out where it came from.

Then, on April 10, the mystery ended: the official @HappyHorseATH account confirmed the model was built by Alibaba's ATH AI Innovation Unit (part of the Taotian Group). Artificial Analysis now lists the model under the label Alibaba-ATH.

How Video Arena Works — And Why This Matters

Artificial Analysis runs a blind comparison arena where users are shown two videos generated by different models from the same prompt, without knowing which model produced which. They vote for the output that's more faithful to the prompt and higher quality, with model identities revealed only after voting. Votes feed into an Elo rating system — the same methodology used in chess rankings.

The Benchmark Numbers

Happy Horse 1.0's dominance spans multiple categories. As of the April 9–11, 2026 snapshot on Artificial Analysis, the model holds the following rankings:

Text-to-Video (no audio): Elo ~1,370, approximately 97 points ahead of second-place Seedance 2.0 from ByteDance. An Elo gap of ~60 points translates to roughly a 58–59% win rate in head-to-head matchups. Happy Horse's margin suggests dominance, not a statistical fluke.

Text-to-Video (with audio): Elo ~1,236, also ranked #1. This is significant — as Gigazine reports, HappyHorse achieves joint audio-video generation in a single pass, ranking first even when audio synchronization is factored in.

Image-to-Video (no audio): Elo ~1,415, an even larger margin over the field.

Note: Elo scores for newly-listed models fluctuate as more votes accumulate. Different sources report slightly different snapshot scores (ranging from 1,333 to 1,388 in T2V no-audio), depending on when the data was captured. The figures above reflect the most commonly cited range as of mid-April 2026.

The Technical Speculation

According to FlowVideo AI's analysis, unofficial sources describe Happy Horse as using a single-stream, 40-layer Transformer architecture with joint denoising across text, image, video, and audio modalities. The reported parameter count is approximately 15 billion, with all modalities sharing the same attention space — no cross-attention between separate encoders.

This architecture differs notably from competitors. Seedance 2.0 uses a dual-branch design with an Attention Bridge for audio-video sync. Kling 3.0 employs cascaded diffusion with separate super-resolution stages. Happy Horse's unified approach appears to excel in pure visual quality, with its audio capability — while still #1 in T2V with audio — showing a smaller lead compared to its dominant no-audio scores.

From Mystery to Confirmation: Alibaba's ATH Unit

The strongest theory — now confirmed — traced Happy Horse to Alibaba's Taotian Group, specifically its Future Life Laboratory operating under the ATH AI Innovation Unit. As Gigazine documented, the model's Chinese-Japanese-Korean language support, anonymous pre-launch testing pattern, and traced connections to Alibaba-linked researchers all pointed in this direction. Chinese labs have form here — both Pony Alpha and GLM-5 were anonymously tested before official reveals.

The anonymous testing appears to have been a deliberate strategy: validate performance claims through third-party benchmarking before a formal launch. With the model now officially attributed to Alibaba-ATH on Artificial Analysis, the strategy worked — the benchmarks spoke before the marketing did.

The Availability Gap

Despite the #1 ranking, Happy Horse 1.0 remains largely inaccessible. No public API, no downloadable weights, no published model card, and no pricing information exist as of April 11, 2026. The gap between a #1 ranking and actual production usability remains significant.

The Scam Problem

Within 48 hours of the ranking, over a dozen fake "official" Happy Horse websites appeared, including domains like happyhorse.app, happy-horse.ai, and happyhorse-ai.com, all offering paid video generation services. Chinese AI community members and Artificial Analysis itself have issued warnings: do not pay for services on any of these sites. Until Alibaba releases an official access channel, any site claiming to offer Happy Horse generation is almost certainly fraudulent.

The Bigger Picture

Happy Horse 1.0's rapid rise highlights a growing trend in AI development: the gap between Chinese and Western AI labs in multimodal generation is closing fast, and in some cases, reversing. While US companies dominate the language model conversation, Chinese labs are quietly building world-class video generation systems.

The anonymous-debut-then-confirm playbook — test in public benchmarks without attribution, let performance speak, then claim ownership — could become the new template for competitive AI launches. For builders and enterprises, the lesson is clear: watch the leaderboards, but don't confuse a ranking signal with a production-ready tool.

Sources: Artificial Analysis Video Arena, FlowVideo AI, Gigazine, Intellectia AI, GrokVideoGenerator