Skip to content

Product1 publisher3 min readPublished

Reaching developers in China without a VPN takes a model hub beyond Hugging Face

Alibaba's ModelScope hosts more than 170,000 models for developers in China, where regulators blocked Hugging Face in 2023. Teams shipping open weights to Chinese users have to plan a domestic mirror with usage figures that come mostly from the hubs themselves.

The Product Desk · Product desk

Illustration accompanying Reaching developers in China without a VPN takes a model hub beyond Hugging Face

What happened

  • OSChina launched a second domestic hub, MoArk, in 2023, and it now serves about 20,000 models.
  • Hugging Face still hosts more than 3 million open models, far more than either Chinese hub.
  • Nvidia, which announced a $12.9 billion deal for Hugging Face in September, flagged in a filing that regulators could ban Chinese models from the site.
  • ModelScope and MoArk both offer model testing and customization, with some free usage and charges for extra computing power and advanced features.

Compiled by The Product DeskSomething wrong?How this is made

Why it matters

  • exposure Chinese labs that publish only on Hugging Face depend on a U.S.-owned site that could be hit by bans the Trump administration has reportedly discussed, and a domestic copy is their hedge.
  • cost Publishers who point Chinese users at a domestic hub for testing or fine-tuning are sending them to a paid service once the free usage runs out.
  • precedent If models follow code, a domestic hub can gain official backing while the foreign one stays the gateway, and publishers keep two copies current for years.
  • decision Maintainers whose installers fetch weights from Hugging Face now have two options for users in China without a VPN: run a second release pipeline or accept that the install fails.

A developer in China runs a setup script that pulls model weights from a Hugging Face address. With no VPN switched on, the download fails. Regulators blocked the site in 2023 and did not give a specific reason [1][2].

Teams outside China tend to tell themselves that Chinese developers get around the block anyway. What users actually do is split. VPN use is common in the tech industry, because developers need blocked sites such as Google and X for their work [10]. Chinese AI labs keep releasing open models on Hugging Face, and the government tolerates them getting around the block to do it [5]. But authorities have occasionally penalized people for getting around the firewall without authorization [11]. Xu Yong, chief executive of OSChina, told Rest of World that not everyone is able to use a virtual private network all the time [6]. His company runs MoArk [4].

The domestic hubs have far smaller catalogs. ModelScope has roughly one model for every 18 on Hugging Face, and MoArk's count is under 1% of Hugging Face's [1][2]. ModelScope said in March that it had 250 million users [18]. That number is self-reported, and the report does not break it down by country or by download activity.

Code distribution in China has already split once. The government briefly blocked GitHub in 2013 and later restored access. In 2020 it endorsed Gitee, a domestic version also owned by OSChina, while GitHub stayed popular as the gateway to the global software community [12]. Xu expects models to move faster. "In the AI era, China is developing an independent ecosystem faster than in the internet era," he said [7].

Rebecca Arcesati, a researcher at the Mercator Institute for China Studies, said Beijing recognizes that access to international open-source platforms matters to its tech industry [16]. She said there is still a real concern in China that access to the U.S.-based platforms could be disrupted at any point [15].

I think teams that count Chinese-speaking developers in their audience should publish a mirror on ModelScope, the larger of the two hubs at about 8.5 times MoArk's catalog [3]. The tradeoff is a second release pipeline. If the two copies drift apart, the support tickets come from users who cannot tell which version they pulled.

Two axes set how urgent that mirror is. One is whether the product fetches weights by itself, at install or at runtime, or whether a person downloads them once by hand. The other is whether target users get around the block routinely, or work somewhere that doing so without authorization is a risk they avoid [11]. If the product fetches weights automatically for users without a reliable VPN, it does not install unless a domestic mirror exists. If the product fetches automatically for VPN users, the mirror is a fallback address in the script. Users who download by hand and have no VPN can only find you on a domestic hub. Users who download by hand and have a VPN are the one group Hugging Face alone reaches today. That access is what Beijing would prefer to replace, in Arcesati's account. "From the Chinese government's perspective, it will be ideal if the entire technology stack for AI, including the software tools and libraries, could be indigenized as much as possible," she said [17].

What to watch

  • Whether U.S. regulators act on the reported discussions of banning Chinese models, and whether any rule reaches Hugging Face once Nvidia owns it.
  • Whether Beijing formally endorses ModelScope or MoArk the way it endorsed Gitee for code in 2020.
  • Independent download data by region showing how many developers in China pull from domestic hubs and how many reach Hugging Face through VPNs.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories