The Anonymous AI Video Model That Beat Sora — Then Revealed Itself as Alibaba
HappyHorse 1.0 appeared on April 10 as an anonymous open-source video model, immediately topped the Artificial Analysis Video Arena above Sora and Kling, then was revealed to be from Alibaba's Taotian Group — a calculated stealth launch that succeeded beyond expectations.
Original sourceOn April 10, 2026, a mysterious video generation model called HappyHorse 1.0 appeared on Hugging Face with minimal documentation and no corporate attribution. Within 48 hours, it had climbed to the top of the Artificial Analysis Video Arena — the most respected blind evaluation platform for video generation AI — defeating OpenAI's Sora 2 Pro, Bytedance's Seedance 2.0, and Kuaishou's Kling 3.0 in side-by-side human preference comparisons.
The model's technical credentials are real: 15 billion parameters, 1080p output, and native audio generation in a single inference pass — a significant capability gap over existing open-source alternatives. The commercial license and fully open weights made it immediately usable for production applications.
Then came the reveal. Investigative reporting by Yicai Global identified the model's origins at Alibaba's Taotian Group (the division running Taobao and Tmall). The anonymous launch was deliberate: rather than risk the model being dismissed as Chinese corporate propaganda or facing geopolitical skepticism, the team released it without attribution and let the benchmark results speak for themselves.
The strategy worked. By the time the corporate origin was revealed, the model had already won substantial credibility in Western AI communities on pure merit. For the local AI video generation community — which has been building on Wan 2.1, CogVideoX, and similar open models — HappyHorse represents a step-change in quality. Self-hostable, commercially licensed, and now #1 on the arena: it's the Stable Diffusion moment for video generation that the community has been waiting for.
The broader implication is about trust architecture in the open-source AI ecosystem. Anonymous releases that let models compete on merit, then reveal provenance after credibility is established, may become a common tactic for labs whose national origin creates headwinds in Western markets.
Panel Takes
The Builder
Developer Perspective
“The technical result stands regardless of the corporate play: open-source video gen just crossed the frontier threshold. Every team building video pipelines should be evaluating HappyHorse immediately. The anonymous launch is a PR strategy, but the model is real.”
The Skeptic
Reality Check
“Anonymous corporate releases are a trust problem regardless of quality. We don't know the training data, the safety evaluations, or the model's behavior on adversarial prompts. 'It scored well on one arena' is not sufficient due diligence for production adoption, especially when the true origin was deliberately obscured.”
The Futurist
Big Picture
“This is what the commoditization of frontier AI looks like: open-weight models that beat closed commercial offerings, released by entities with competitive reasons to obscure their identity. Video generation is now democratized. The next question is what comes after the current leaderboard paradigm when every model scores in the same band.”