Alibaba Tops Video‑Gen Benchmarks

A mystery video model called HappyHorse-1.0 appeared on leaderboards this week with no company name attached, then Alibaba confirmed on April 10 that the model came from its ATH AI Innovation Unit and is still in internal testing. (cnbc.com) The surprise was not just that Alibaba built it. The surprise was that HappyHorse went straight to No. 1 on Artificial Analysis’s blind-vote text-to-video leaderboard, where people compare clips without seeing which company made them. (artificialanalysis.ai) On that leaderboard, HappyHorse-1.0 was listed at 1,355 Elo points on April 10, ahead of ByteDance’s Dreamina Seedance 2.0 at 1,273 and ahead of OpenAI’s Sora 2 Pro at 1,196. Elo is the same rating logic used in chess: beat strong opponents in head-to-head matchups and your score rises. (artificialanalysis.ai) Video generation is the branch of artificial intelligence that turns a written prompt or a still image into a moving clip. The hard part is not making one pretty frame, but keeping the same person, object, and camera motion coherent across dozens of frames in a row. (openaccess.thecvf.com) That is why the industry leans on benchmarks like VBench, a research suite presented at the 2024 Conference on Computer Vision and Pattern Recognition that breaks video quality into dimensions like subject consistency, motion, and aesthetics. It is basically a report card that checks whether a model keeps the story straight while the frames keep moving. (openaccess.thecvf.com) China’s internet companies have been sprinting in this category because short video, advertising, and live commerce are already giant businesses there. A better model can cut the cost of making product demos, animated ads, and social clips the way a faster camera crew cuts shooting days. (marketwatch.com) Alibaba was already in this race before HappyHorse. Its Wan family was open-sourced earlier, and Alibaba Cloud said on April 7 that the new Wan2.7-Video release was built to improve delivery quality and creative efficiency for creators and businesses. (github.com, alibabacloud.com) That makes the stealth launch more revealing than a normal product update. Alibaba was testing whether raw output quality could win attention on its own, and the answer was yes: the model climbed before the company took credit for it. (bloomberg.com, theinformation.com) The other company watching this most closely is ByteDance, because ByteDance’s Dreamina Seedance 2.0 had been the model to beat on the same leaderboard. When one Chinese company jumps another in a benchmark that creators actually watch, it shifts where developers, cloud customers, and investors think momentum is building. (artificialanalysis.ai, marketwatch.com) Alibaba said API access is coming soon, which is the practical next step. A benchmark win gets headlines, but an application programming interface is what lets ad tools, editing apps, and commerce platforms plug the model into real production pipelines. (technode.com) If HappyHorse keeps its lead once outsiders can test it at scale, the bigger shift is not one more flashy demo. It is that video generation may start moving from a novelty clip maker into ordinary software for product pages, brand campaigns, and short-form entertainment, with Alibaba trying to supply both the model and the cloud behind it. (cnbc.com, alibabacloud.com)