周三，Artificial Analysis 在其 text-to-video 基准测试中将一个新模型列为第一名；阿里巴巴于周五确认该模型就是 Happy Horse。这个模型在几周近乎沉默后出现，且在阿里巴巴先前围绕 Qwen 应用与之前 Wan 影片更新所进行的大量宣传之后浮现。由于今年是中国农历马年，外界曾推测可能是 DeepSeek 或 ByteDance 的 Seedance 2.0，因此确认榜首来自阿里巴巴本身是一个明显反转。
内部追溯显示，Happy Horse 由 Alibaba Token Hub 的 ATH AI Innovation Unit 制作，而不是打造 Wan 的 Tongyi Lab，表明阿里巴巴正在进行多条平行研发，类似中国旧有的「race-horse」模式，即多个团队争夺同一类别。Xiaomi 的 Hunter Alpha 曾以匿名方式先出现在基准榜后再被认领，且 Xiaomi 在身份确认后立即向开发者开放；相比之下，阿里巴巴将在本月稍后才全面开放开发者使用。基准竞争还受到 OpenAI 在 3 月终止 Sora、将资源转向 AI agents 的影响，凸显视讯模型的高固定与营运成本。
文章强调结构性压力：中国的人才与算力均偏紧缺，先进 Nvidia 晶片受限，但电力成本较低，故 AI 部署经济仍可成立。视讯 AI 具有吸引力，因为中国消费者仍对软体支出较谨慎，营收转化存在不确定性但也可能很有价值。阿里巴巴在 Junyang Lin 离职及组织重组后，改以 ATH 面向商务客户，并在短期内连续推出至少三个闭源模型，显示其偏离过去开源重心。公司承诺投资逾 $53 billion，且 2026 年营收转化备受监督；在投资者耐心下滑情况下，管理层被期望在维持模型领先的同时，不断调整商业模式。
On Wednesday, Artificial Analysis placed a new model at the top of its text-to-video benchmark, and Alibaba confirmed on Friday that it was Happy Horse. The model appeared after weeks of near silence after Alibaba’s earlier publicity around the Qwen app and previous Wan video update. Because this was the Chinese Lunar Year of the Horse, observers had speculated about DeepSeek or ByteDance’s Seedance 2.0, so the confirmation that the top model came from Alibaba was a notable reversal.
Internal sources show that Happy Horse was built by Alibaba Token Hub’s ATH AI Innovation Unit, not Tongyi Lab, which built Wan. This indicates Alibaba is running parallel development tracks, similar to the older Chinese 'race-horse' strategy where multiple teams compete in the same category. Xiaomi’s Hunter Alpha earlier appeared anonymously at the top of a benchmark before its identity was claimed; Xiaomi then opened it quickly to developers. Alibaba, in contrast, said broad developer access will start later this month. The benchmark environment is also affected by OpenAI’s March decision to discontinue Sora and shift to AI agents, highlighting high fixed and operating costs for video models. (Key numbers: 3)
The article frames structural pressure: in China, talent and compute are still tight and advanced Nvidia chips are constrained, while cheaper electricity helps keep AI deployment economics viable. Video AI is attractive because Chinese consumers remain cautious on software spending, making monetization uncertain but potentially important. After Junyang Lin’s departure and restructuring toward ATH for business clients, Alibaba rapidly released at least three closed models, signaling a move away from its earlier open-source emphasis. With promised AI investment of more than $53 billion and revenue conversion into 2026 under scrutiny, management teams are expected to defend model leadership while continually resetting business models as investor patience shrinks.