HappyHorse and Seedance large models Photo: VCG

An artificial intelligence (AI) video creation platform HappyHorse-1.0 has made itself a "black horse" lately. The AI model, with multiple clues indicating it is most possibly developed by a Chinese team affiliated with Alibaba, has topped both text-to-video and image-to-video leaderboards of independent third-party AI benchmarking platform Artificial Analysis during the past week, generating a buzz across the AI industry and the capital market.

It was just two months after Seedance 2.0, an AI video model developed by Chinese technology giant ByteDance, the parent company of TikTok, rocked both Silicon Valley and Hollywood with its cinema-quality video generation in February. The AI model now ranks second in Artificial Analysis' leaderboard of text-to-video and image-to-video rankings.

The rapid rise of those Chinese AI models also reminds industry practitioners of the "DeepSeek moment" at the beginning of 2025, when the Chinese model suddenly stunned the global AI community with its extraordinary performance at a fraction of the cost of leading Western models. However, things are quite different one year on, as China has seen a steady stream of "black horse" AI models - so much so that these rapid breakthroughs have become almost expected rather than surprising.

"This trend is not a coincidence, but a new norm driven by multiple factors, including Chinese companies' high performance-per-compute efficiency, the deep integration of algorithmic engineering for efficiency improvement with real-world commercial scenarios, as well as a strategic push toward open-source development," Tian Feng, former dean of SenseTime's Intelligence Industry Research Institute, told the Global Times on Thursday.


Topping the list

In the text-to-video leaderboard, HappyHorse-1.0 achieved an Elo score of 1,383, outperforming Seedance 2.0 by 111 points. Artificial Analysis uses a blind testing mechanism for its leaderboard. Users compare and select between generated videos without knowing the model source, and the results are aggregated using an Elo rating system similar to that used in chess, which creates a ranking based on real user preferences, according to a report by news portal thepaper.cn.

Early community test examples show that Happy Horse can easily handle complex dynamic scenes, such as a long-duration video showing "flowers in the same vase blooming and withering over two weeks" with coherent visuals and realistic lighting, which exceeds the usual performance of similar models, according to Chinese media reports. 

According to an introduction on the model's website, HappyHorse 1.0 is reportedly developed by the Future Life Lab team of Taotian Group (Alibaba), led by Zhang Di — a veteran of China's top AI video labs. Alibaba has not responded to a request for comment from the Global Times.

While HappyHorse-1.0 topped the text-to-video ranking of Artificial Analysis' leaderboard, the remaining top four AI models are Seedance 2.0, SkyReels V4, Kling 3.0 1080p (Pro) and Kling 3.0 Omni 1080p (Pro)- all developed by Chinese companies. The Grok-image-video model, developed by xAI, ranks sixth on the leaderboard.

"Unlike the traditional approach that overly relies on Scaling Law, the HappyHorse model - through architectural and algorithmic innovations - achieves 8-step ultra-fast inference with an extremely small 15-billion parameter model," Tian said, stressing that this deep optimization of algorithmic efficiency is the core technological advantage that enables the model to outperform its rivals. 

The scaling law refers to the observed phenomenon that the performance of AI models tends to improve predictably and reliably as the model size (parameters), training data volume, and computational resources are scaled up. This principle has become a core development philosophy embraced by leading US AI companies, including OpenAI, Google DeepMind, Anthropic, and xAI.

In addition, large digital industries such as e-commerce and film in China have provided Chinese AI models with the world's most demanding and diverse high-quality scenario data, helping accelerate their independent innovation, Chinese industry analysts noted. 

"Instead of being confined to laboratory benchmarking, these models are now continuously refined and iterated through real-world applications — including virtual try-ons, film pre-visualization, and dynamic physical interactions," Tian said.

Fierce competition
 
In March, OpenAI announced the shutdown of its AI video generation model Sora, according to a BBC report, marking the formal exit of what was once regarded as the benchmark product in AI video generation. 

"The shutdown of Sora stems from prohibitively high computing costs, copyright compliance pressures, and commercialization difficulties. It also serves as a wake-up call for the global multimodal AI race: the era of pure technological showcasing is over. Practicality, controllability, and real-world applicability have become the core competitiveness of AI models," Cheng Yin, the AI research manager at IDC China, said in a research note sent to the Global Times. 

As global competition intensifies, independent technological architectures, compliant and secure data systems, and efficient, inclusive industrial value will be the key to competition in the next stage, according to Cheng. 

Tian believed the next stage of global competition will be a cut-throat "life-or-death" test based on a number of indexes, including product-market fit (PMF), ecosystem and technical engineering efficiency.

"In the PMF regard, China's rising 'black horses' adopt a very pragmatic approach compared with foreign rivals. For example, ByteDance's Seedance 2.0 has outperformed rivals in professional multi-shot cinematic applications for film and TV, while HappyHorse could directly address key pain points in e-commerce. Both have strategically focused on high-value B2B scenarios and highly sticky use cases that can quickly generate tangible revenues," Tian explained.

Industry analysts noted that an open-source strategy will accelerate the development of a thriving ecosystem. Such a strategy will also lower inference and deployment costs, while also reducing the barrier to entry, enabling small and medium-sized enterprises, startup ventures, and individual developers to use, fine-tune, and commercialize these models at scale with unprecedented speed.

According to the official website, HappyHorse-1.0 is designed to be open source, and technical details will come with the open-source release, though GitHub and HuggingFace links are not available yet.