When Beijing blocked Hugging Face in 2023, it did something the government usually does: it accidentally created a market. The ban wasn't about the models — Chinese AI companies were shipping powerful open models onto that very platform for the world to download. It was about the pipe: a U.S.-based hub hosting the shared infrastructure of global AI, reachable only through a VPN, disruptable at any moment by a regulator with an axe. So the domestic answer started assembling itself — Alibaba's ModelScope on one side, OSChina's MoArk on the other — and now, eighteen months later, the race to become "China's Hugging Face" has real numbers, real users, and a geopolitical price tag attached to it: Nvidia just paid $12.9 billion to buy the original. The two ecosystems are no longer just competitors. They're the two poles of the same weapon.
Rest of World's reporting on the two platforms is the clearest look yet at how the open-source world is splitting into national halves — and why the split is more asymmetric than anyone on either side of the Strait is comfortable admitting.
The Two Contenders
ModelScope, built by Alibaba, launched in 2022 and now hosts more than 170,000 models with 250 million users, by its own March reporting. The growth strategy is community-first: university hackathons where students compete to build AI applications, and a physical coworking and event space in Hangzhou for AI entrepreneurs. It's the platform that looks most like Hugging Face — models, datasets, testing, customization, some free usage, paid compute on top.
MoArk, from OSChina, launched in 2023 and serves roughly 20,000 commonly used models. But its most important team isn't in product. It's the group of engineers deployed specifically to adapt open models to run on the different kinds of Chinese-made chips — Huawei and the rest of the domestic accelerator race — because, in the words of OSChina chief executive Xu Yong, "It's our mission to make sure mainstream open models can run on Chinese chips." That's not a hosting feature. That's the integration layer of a national compute strategy.
Against them sits the benchmark: Hugging Face, with more than 3 million open models. The gap is not close. It's three orders of magnitude in library, one of orders of magnitude in momentum — and it's the number that tells you exactly how much time China's domestic open ecosystem still has to catch up before the U.S. side consolidates the lead entirely.
The Paradox: One Pipe, Two Directions
Here's the part that makes this story genuinely strange, and worth understanding. The Chinese government blocked Hugging Face for its own domestic users — because the stack must be indigenized, as Rebecca Arcesati of the Mercator Institute for China Studies put it: "the entire technology stack for AI, including the software tools and libraries, [should be] indigenized as much as possible." At the same time, Beijing quietly tolerates Chinese AI labs using that same blocked platform to share Chinese open models with the world. The block runs one way — domestic users in, foreign infrastructure out of the country's reach — while the export of Chinese models runs the other way, past the block, on purpose. One regulator, two directions, two different strategic goals. The pipeline is nationalized; the products are global.
That's the same architecture as the chip war itself: export controls are a wall with a door in it, and the door only opens in one direction.
The U.S. Side of the Split
The mirror-image pressure is building fast. The Trump administration, keen to hold America's lead in AI, has reportedly discussed bans on Chinese open models — the very models that flood Hugging Face and that the global open ecosystem now runs on. And Nvidia's own SEC filing on the $12.9 billion Hugging Face acquisition flagged the risk that regulators could ban Chinese models from being shared on the site it just bought. Read that sentence slowly: the company that just spent $12.9 billion to control the open-model hub is formally warning its investors that the host nation's regulators might order it to expel the largest bloc of open-model contributors. The acquisition is no longer just an ecosystem play. It's a geopolitical asset with a compliance footnote.
Pair that with the compute context — China's AI industry is building under a chip-export ban, racing to design its own accelerators — and you have the full picture: two countries are each nationalizing half of the open stack, and the "open" in "open-source" is quietly becoming "open-within-your-fence."
What the Developers Actually Say
The floor of the story is one sentence from Chen Yunfei, an independent AI developer in Kunming: "Hugging Face and GitHub are the default in the open-source space. They offer a greater variety. Domestic platforms need to convince us why we should use them." The main attraction of ModelScope and MoArk, by developers' own accounting, is download speed — a practical win that only matters when the Great Firewall is the bottleneck. The strategic gap between "fast to download" and "I build the default here" is where the next few years of China's open ecosystem will either be won or lost. OSChina's chief put it in one line: "In the AI era, China is developing an independent ecosystem faster than in the internet era." The internet era took a decade to rebuild; the AI era is getting its answer in a few years.
What It Means
1. The 3M-vs-190K model gap is the real headline, and it's a countdown. Hugging Face's library is roughly fifteen times the combined total of China's two platforms. But libraries aren't static: they compound. Every model MoArk or ModelScope hosts and adapts to domestic chips is a data point toward a self-sustaining national open ecosystem — and the compounding starts when a critical mass of the most-used open models can be found and fine-tuned without a VPN. The countdown is to that critical mass, not to today's numbers.
2. MoArk's chip-adaptation team is the story inside the story. A hosting platform that also maintains a team to port every mainstream open model onto Chinese-made accelerators is not competing on models. It's competing on the integration layer of the national compute strategy — the same layer Huawei and DeepSeek have been racing to own. Whoever holds "the model runs on your chip, for free" holds leverage over every Chinese inference business that has to buy chips from the domestic stack. That's a platform business with a government subsidy hidden in the engine room.
3. The "open with a fence" era has arrived, and it changes the value of every open model. An open model's worth is no longer just its weights. It's which fence it lives in, which platform can host and fine-tune it, and which regulator can order its removal. Nvidia's own filing is the first formal document admitting that the global open hub is now a controlled asset. That reframes the $12.9 billion not as an acquisition price but as a valuation of control — and it tells you what the other side is doing the same calculation for.
4. The developers' loyalty is the swing variable. The whole edifice — national platforms, national chips, national fences — is load-bearing only if the people writing the code actually live inside it. As long as Chinese developers default to Hugging Face and GitHub "with a VPN," the domestic platforms are fast but secondary. The moment the default flips, the 3M library stops mattering and the 190K becomes the whole world. Watch the hackathon attendance, the Hangzhou coworking occupancy, and the share of new Chinese model releases that land on ModelScope first. Those are the leading indicators of the flip.
🔥 Hot Takes
1. Nvidia's $12.9B Hugging Face bid is now a national-security purchase, and both sides know it. The SEC filing's "regulators could ban Chinese models from the site" line is the first time a U.S. acquisition of an open-source hub has had to disclose that its core inventory might be illegal to host. That's not a footnote. That's the official record now saying: the open model you downloaded might be contraband depending on which government you answer to. The "open" in open-source just became the most contested word in tech policy.
2. "170,000 models" is the number that flatters ModelScope — the 250 million users is the number that matters. A model library that nobody else on the continent can access without a VPN, with 250 million registered users, is the equivalent of the U.S. building an interstate highway system that only domestic trucks can legally use. You can measure it in models hosted, or in the fact that it's now the only open highway in the country. The second measurement is the one that compounds.
3. The real product neither ModelScope nor MoArk is selling is the absence of risk. What they're actually selling to a Chinese developer is: your work won't be yanked by a foreign regulator, your downloads won't be throttled by a foreign pipe, and your fine-tunes won't be subject to a foreign terms-of-service change. That's a reliability product with a sovereignty wrapper — and it's exactly the product that makes the "indigenized stack" thesis sellable to a developer who would otherwise just use a VPN. If they can't sell that, the platforms stay fast-but-secondary forever.
The Bottom Line
China's Hugging Face is not one company. It's two: Alibaba's community machine and OSChina's chip-adaptation engine, racing the same countdown against a U.S. hub that just got bought for $12.9 billion and is now formally flagged as a regulatory minefield. The open-source world split into national fences, and the fence China is building has a genuine consumer product inside it — faster, sovereign, and compounding. The question isn't whether it catches up on model count. It's whether the developers' default flips before the other side finishes fortifying its own. That flip, when it comes, will be the moment "open-source" officially stopped meaning "anyone" and started meaning "everyone on our side." Both sides are already pricing it in.