Chinese video models take eight of ten seats on a public AI leaderboard

Chinese video-generation models occupy eight of the top ten places on Artificial Analysis's text-to-video leaderboard with audio, the South China Morning Post reported on Saturday. Alibaba's Wan 3.0 sits first, ahead of Google Gemini Omni Flash. A MiniMax H3 checkpoint post-trained by the US platform fal is third. The original open-weight H3 and ByteDance's Seedance 2.0 follow. SCMP is owned by Alibaba. That ownership does not invent the ranking. It does mean the first account of the ranking comes from a paper inside the group that sells Wan.

The same models lead several related boards for image-to-video and video editing in blind user tests with identical prompts, the paper said. Analysts quoted in the piece pointed at three advantages that have little to do with raw training compute: a huge short-video market, copyright rules that give generators more room than Hollywood contracts allow, and prices set to win volume. OpenAI and Anthropic still lead the closed large-language-model race on most public scores. Video is a different contest. It rewards data from Douyin, Kuaishou and Bilibili more than it rewards another cluster of GPUs in Oregon.

The commercial proof is already on Chinese timelines. State media told CNN that more than 95 percent of microdramas released in the first quarter of 2026 were made with AI. Global Times, citing industry counts, put first-quarter releases at about 128,000 titles. Zhu Xinmei of the National Radio and Television Administration's development centre called that shift a change in the productive forces of broadcast and television. The language is official. The volume is not.

One title made the argument in public last week. The 26-minute film Yaoyi Bu: Wen Cangsheng, The Book of Strange Tales: Asking about the People, passed 14.5 million views on Bilibili. Creator Qingguadan used ByteDance Seedance 2.5, OpenAI's GPT Image tool and Jimeng AI. The plot follows a historian sent to catalogue monsters who finds famine and forced labour instead, then files the victims as monsters to keep them alive. Viewers argued about the moral and barely mentioned the renderer. That is the market test Chinese studios wanted. Half of a Youtiao, another AI short, had already cleared 1.1 million likes in July on memory rather than on novelty.

Hiring followed the models. CCTV said demand for AI-assisted video editors jumped 179 percent year on year in June. Monthly pay for those roles runs from 8,000 to 20,000 yuan. The same report put the talent gap above 500,000 people. Feng Anrong, a former screenwriter who now runs an AI studio, told Global Times that overseas platforms are already calling Chinese weights. Fu, another producer in that piece, said Chinese short-drama formats have moved into Japan, Korea, Southeast Asia, the Middle East and North America.

None of this settles the language-model race. It does settle a narrower claim that was fashionable a year ago: that video generation would stay an American luxury product. Wan, H3 and Seedance are cheap enough for a microdrama mill and good enough to hold a character's face for 26 minutes. Union rules in Los Angeles still treat that combination as a threat. Chinese regulators treat it as output. The leaderboard is the scoreboard those two systems now share. Eight of ten is the number a reader should take from Saturday's report, with the ownership of the messenger attached.