China’s Open-Model Platforms Test a Local Alternative to Hugging Face

ModelScope and MoArk are trying to turn China’s restricted access to Hugging Face into a domestic market for open-model infrastructure, Rest of World reported.

The contest sits inside Beijing’s larger AI trade-off.

Regulators blocked the U.S.-based Hugging Face platform in 2023, limiting ordinary domestic access to a global hub where developers share models and data sets.

At the same time, China’s leading AI labs still need channels to release Chinese models internationally and to track work published outside the country.

That gap has given Chinese alternatives a clearer role.

Alibaba launched ModelScope in 2022, and the platform now hosts more than 170,000 models.

OSChina followed in 2023 with MoArk, which serves about 20,000 commonly used models.

Both platforms offer model testing and customization, provide some free usage and charge for additional computing power or advanced features.

The scale difference remains large.

Hugging Face hosts more than 3 million open models, making it the default global marketplace for many developers.

ModelScope said in March that it had 170,000 models and 250 million users, while MoArk is still far smaller by model count.

That leaves the Chinese services competing less as direct replacements today than as infrastructure designed around local access, language needs and regulatory durability.

OSChina chief executive Xu Yong framed the domestic push around the same firewall that shaped China’s earlier internet economy.

Not every developer can rely on a VPN all the time, and a Chinese-language ecosystem can make open models easier to reach inside the country.

His view also places MoArk inside a broader effort to make AI tools less dependent on foreign platforms.

The market structure is shaped by security concerns on both sides.

Chinese developers have long used international open-source platforms to collaborate and to reach a global audience.

GitHub remained important even after China briefly blocked it in 2013 and later endorsed OSChina’s Gitee as a domestic code-hosting alternative.

For AI, the government’s tolerance is similarly uneven: international access supports industrial development, but domestic platforms are viewed as more controllable and less exposed to U.S. regulation.

Rebecca Arcesati, a researcher at the Mercator Institute for China Studies, described that exposure as a live concern for China.

The Trump administration has reportedly discussed bans on Chinese models, and Nvidia’s regulatory filing on its planned $12.9 billion acquisition of Hugging Face flagged the possibility that regulators could bar Chinese models from being shared on the site.

Computing limits add another layer to the domestic-platform race.

U.S. controls have restricted sales of top AI chips to China, pushing local companies to develop alternatives.

MoArk has assigned engineers to adapt AI models for different Chinese chips, making hardware compatibility part of its pitch rather than just hosting and discovery.

ModelScope has tried to build developer habits through university hackathons and a Hangzhou coworking and event space for AI entrepreneurs.

MoArk is leaning on adaptation work and OSChina’s open-source developer base.

Both approaches aim to make the Chinese platforms part of daily workflows before a future access shock makes foreign hubs harder to use.

Developers have not abandoned Hugging Face.

Chen Yunfei, an independent AI developer in Kunming, said domestic platforms offer faster downloads but still need to show why developers should choose them over Hugging Face and GitHub, which remain the defaults for variety and global open-source reach.