China‘s AI sector is developing an independent ecosystem, with several open-source platforms emerging as domestic alternatives to Hugging Face, a US-based open-source hub blocked by Chinese regulators in 2023. Xu Yong, chief executive of OSChina, notes that China needs to build a self-reliant AI ecosystem to serve the Chinese-speaking population.
Alibaba, founded in 1999 by Chinese entrepreneur Jack Ma, launched a Hugging Face-like open-model platform, ModelScope, in 2022, which now hosts more than 170,000 models. OSChina launched another open-model site, MoArk, in 2023, serving some 20,000 models.
Domestic Alternatives Gain Traction
According to Xu Yong, not everyone can use a virtual private network (VPN) all the time, highlighting the need for domestic alternatives. Xu believes that China is developing an independent ecosystem faster than in the internet era.
Open models, unlike closed models from OpenAI and Anthropic, can be downloaded, fine-tuned, and deployed on local hardware. This has led to a surge in interest in domestic open-source platforms, with ModelScope and MoArk offering services similar to Hugging Face, including model testing and customization.
Rebecca Arcesati, a researcher with the think tank Mercator Institute for China Studies, notes that there is still a real concern in China that access to US-based platforms could be disrupted at any point. This concern has driven the development of domestic platforms, which are seen as more secure and reliable.
Chinese AI companies have been releasing powerful open models on Hugging Face, which can be downloaded and deployed across the world. However, the government also believes that domestic platforms are essential for the development of China’s tech industry, and has endorsed the use of domestic alternatives.
Challenges Ahead
Despite the growth of domestic platforms, they still fall far behind Hugging Face, which hosts more than 3 million open models. ModelScope has 170,000 models and 250 million users, while MoArk hosts more than 20,000 commonly used models.
To attract users, ModelScope has hosted university hackathons and opened a coworking and event space in Hangzhou for AI entrepreneurs. MoArk has deployed a team of engineers to adapt AI models to run on different types of Chinese chips, establishing itself as part of China’s own AI ecosystem.
Chen Yunfei, an independent AI developer in Kunming, notes that the main attraction of domestic platforms is faster download speeds, though they still prefer accessing Hugging Face with a VPN. The domestic platforms need to convince developers why they should use them, rather than relying on US-based platforms.
The Chinese AI industry has faced a shortage of computing power, as the US has banned sales of top AI chips to China. Several local companies are now racing to build advanced chips, with MoArk working to adapt AI models to run on Chinese chips.
Regulatory Balance and International Links
Beijing walks a fine line between isolating domestic internet traffic and preserving access for top AI researchers to global breakthroughs. Officials block foreign sites deemed harmful while allowing certain labs to route models abroad through unofficial channels.
Domestic Platforms and Chip Development
Gitee, the home-grown counterpart to a well-known global code-hosting service, received official endorsement in 2020 and now serves as a primary gateway for many programmers. Despite its rise, the original platform remains indispensable for developers seeking the widest range of resources.
Chinese programmers have historically used overseas services to preserve content censored at home, including articles about pandemic restrictions. A brief interruption of the foreign code platform in 2013 sparked widespread protest among the tech community, leading to its swift restoration.
MoArk’s engineering team focuses on optimizing open models for these new Chinese chips, ensuring that widely used algorithms can execute efficiently without reliance on imported hardware. The initiative reflects a broader mission to align the software stack with homegrown silicon.
ModelScope provides a tiered service model, offering basic access at no charge while monetizing additional compute cycles and premium features. This structure mirrors pricing strategies seen on overseas platforms, encouraging users to scale up as project demands increase.
Nvidia’s purchase of the leading U.S. open-source hub for $12.9 billion signals heightened commercial interest in the open-model market. The deal also raised concerns in Washington about potential restrictions on Chinese contributions to the shared repository.