Chinese Video Models Surge After OpenAI’s Sora Shutdown; HappyHorse (欢乐马) Tops AI Rankings

Surprise rival to Sora

It has been reported that just 13 days after OpenAI announced the shutdown of its video model Sora, a new domestic model called HappyHorse (欢乐马) emerged and quickly topped the charts on Artificial Analysis with a score of 1,300. Artificial Analysis is an influential AI-evaluation platform that ranks models via human blind A/B tests coupled with an Elo-style scoring system; users view two clips generated from the same prompt and vote on which feels better, and scores update in real time. Reportedly, the current top-15 of the platform’s video arena is dominated by Chinese models such as ByteDance’s (字节跳动) Seedance 2.0, Kunlun Wanwei’s (昆仑万维) SkyReels V4 and Kuaishou’s (快手) Keling (可灵) 3.0, while overseas entrants like Google’s Veo occupy only a few slots.

Blind tests favour human perception over raw scale

Industry observers say these results reflect that human preference in video generation often tracks qualities like character consistency and temporal stability rather than sheer parameter count. Guosheng Securities has cited sources noting that randomness in generative video—popularly called “抽卡” (gacha draws)—means users typically run the same prompt multiple times (sometimes seven draws for one good shot) before getting a satisfactory clip. Reportedly, HappyHorse—despite being a 15-billion-parameter model, smaller than some Western “parameter giants”—offers native audio-video joint generation and lip-sync in seven languages (including Mandarin, Cantonese, English, Japanese and Korean) with a low word-error rate relative to comparable open models.

Platform integration and creator ecosystems matter

The rise of Seedance 2.0 and Kuaishou’s Keling underlines another dynamic: Chinese content platforms that combine large foundation models with vertical video tools enjoy system-level advantages. ByteDance (字节跳动) and Kuaishou (快手) are not pure model shops; they are creator platforms that can convert model gains into production workflows, distribution and monetization. Analysts quoted by the coverage say that having a generalist large model helps, but true competitive edge comes from deep integration between general models, specialized video models and creator tooling. Amid heightened US–China competition over AI capabilities and chip supply, does this momentum signal an inflection point for China’s video-AI stack? For now, the blind-vote leaderboard suggests Chinese models are resonating with human viewers in ways that raw parameter counts alone do not predict.