A dark horse recently dominated the global AI video evaluation platform. In the early hours of April 8, a video generation model codenamed HappyHorse-1.0 unexpectedly claimed the top spot on Artificial Analysis's Video Arena leaderboard. It ranked first in both text-to-video and image-to-video generation, dethroning ByteDance's Seedance2.0. The sudden appearance sparked widespread speculation about the developer behind this "Happy Horse." The mystery was soon resolved.
On April 10, sources revealed that HappyHorse was developed by Zheng Bo's team under Alibaba's ATH division. Alibaba confirmed that HappyHorse is currently in internal testing, with API access to be opened soon. Following the announcement, investors began buying shares, pushing Alibaba's stock price into positive territory. The market reaction is understandable given HappyHorse's impressive performance. On the Artificial Analysis leaderboard, HappyHorse scored 1,365 points, surpassing Seedance 2.0's 1,273 points. Kuaishou's Kling models ranked fourth and sixth respectively. In image-to-video generation, HappyHorse led the second-place model by 48 points, while the total score difference between second and tenth place was only about 50 points. Artificial Analysis employs a blind testing mechanism where thousands of users evaluate videos generated from the same prompt without knowing the model's origin, making brand bias and ranking manipulation ineffective. HappyHorse's achievement reflects genuine user preference.
The background of this model's development is particularly noteworthy. Zheng Bo is familiar to those following Alibaba's AI developments, though his previous work focused on search, recommendation, and advertising technologies rather than model training. Since joining Alibaba in 2017, he has served as head of Taobao's search and recommendation algorithms, CTO of Alimama, and head of algorithm technology for Taobao and Tmall Group. His team originally belonged to the "Future Life Laboratory" under Taobao and Tmall, which was recently integrated into the ATH division's AI Innovation Business Unit during Alibaba's latest organizational restructuring. This origin story is significant.
Previously, Alibaba's video generation model development was primarily led by the Wanxiang team under the Tongyi Laboratory, representing the company's core AI research division focused on foundational model development. HappyHorse's emergence indicates that the ATH division has cultivated a second team capable of top-tier multimodal model training, with inherent understanding of commercial applications and user needs. This dual-track approach features one laboratory conducting fundamental research while another team drives application innovation from business scenarios, representing a strategic dual-engine structure in multimodal AI rather than simple internal competition.
The timing of HappyHorse's ascent is equally intriguing. In early March, former Qwen head Lin Junyang announced his departure, followed by rumors about other key personnel leaving or transferring positions. This sparked concerns about Alibaba's competitiveness in AI models. However, Alibaba's rapid adjustments became evident: Lin departed in early March, the ATH division was established on March 16, Qwen 3.6 Plus topped global rankings with 1.4 trillion tokens processed on OpenRouter on April 2, and HappyHorse claimed the top video model spot on April 8. Within one month, Alibaba delivered breakthroughs in both language and video models. While Lin's contributions were undeniable, these developments demonstrate that with sufficient technical accumulation and talent depth, individual personnel changes don't undermine organizational capability.
HappyHorse's significance extends beyond proving Alibaba's continued competitiveness. Examining it within Alibaba's broader AI strategy reveals multimodal AI's increasing importance. On April 8, Tongyi Laboratory was upgraded to the Tongyi Large Model Business Unit with Zhou Jingren appointed Chief AI Architect, coinciding with HappyHorse's leaderboard achievement. This timing suggests strategic coordination rather than coincidence. Alibaba needs to signal that its multimodal initiatives represent organized, multi-pronged advancement rather than experimental projects. The Tongyi Large Model Business Unit oversees foundational model development, while the AI Innovation Business Unit approaches multimodal model training from application perspectives. Both units belong to ATH, sharing computing resources and data infrastructure while maintaining differentiated product directions.
Crucially, Alibaba revealed that HappyHorse-1.0 is just one of Zheng Bo's team's multimodal models, with another distinct model scheduled for release soon. The ATH Innovation Business Unit has launched an "AI Era New Interaction Method Exploration Program," with HappyHorse being part of this initiative. This indicates systematic investment in multimodal AI where video generation serves as an entry point toward video understanding, multimodal agents, and new human-computer interactions—any of which could spawn next-generation killer AI applications.
From an industry perspective, HappyHorse's success rewrites established narratives. Previously, AI video generation competition appeared clearly structured: ByteDance's Seedance series leading, Kuaishou's Kling following, and OpenAI's Sora looming overseas. While Alibaba had Wanxiang, its presence seemed limited. HappyHorse shattered this consensus. A team growing from Taobao's ecosystem anonymously submitted a model for blind testing without pre-launch hype and topped the rankings—this approach itself conveys an attitude of relying purely on product capability rather than brand reputation. Notably, users rated HappyHorse superior to Seedance 2.0 in multi-shot coordination, physical motion simulation, and audio-visual synchronization—precisely the capabilities critical for practical applications in film production, advertising creativity, and e-commerce content creation. Zheng Bo's team's understanding of user needs from search recommendation and e-commerce experience likely contributed to this product strength.
For ByteDance, HappyHorse's emergence means Seedance no longer holds undisputed dominance. Once the perception of ceiling-breaking capability shifts, competitive dynamics will be redefined. For the industry, HappyHorse demonstrates that in multimodal model races, traditional underdogs can achieve breakthroughs through diverse technological approaches. This may be the true purpose behind CEO Wu Yongming's recent organizational adjustments—transitioning Alibaba's AI capability from hero-driven to system-driven development. The happy horse has reached the summit, but the real story is just beginning.
Comments