Hugging Face is unreliable from China. GitHub blocked millions of Chinese developers in 2025. China's response — ModelScope, Gitee, PaddlePaddle — has built a parallel open-source ecosystem with 170,000+ models and 25M+ developers in just a few years. The global AI community is splitting along infrastructure lines, not by choice but by Internet fracture. Here's what that means for developers and enterprises everywhere.
For years, the AI community operated on shared infrastructure. Hugging Face was the hub for model distribution. GitHub was the source for code collaboration. PyPI and npm served package distribution globally. The assumption was that open-source AI was borderless.
That assumption is breaking down. In 2025, GitHub blocked access for developers in China and several other countries in response to export control regulations. Hugging Face has become increasingly unreliable from mainland China, with intermittent access and slow download speeds. The infrastructure that the global AI community built together is no longer equally accessible to everyone.
China's response has been rapid and systematic. ModelScope, Alibaba's open-source model platform, now hosts over 170,000 models. Gitee, China's GitHub alternative, has over 25 million registered developers. Baidu's PaddlePaddle ecosystem, originally a deep learning framework, has evolved into a comprehensive AI development platform with model hubs, training tools, and deployment pipelines that operate entirely within China's internet infrastructure.
What makes this significant is not just the numbers — it's the self-sufficiency. Chinese developers can now go from model discovery to deployment without touching any Western infrastructure. The model ecosystem on ModelScope includes not just Chinese models like Qwen, DeepSeek, and Yi, but also re-hosted versions of Llama, Mistral, and other Western models. The training tools on PaddlePaddle support both Chinese and Western hardware. The deployment infrastructure runs on domestic cloud providers.
This parallel ecosystem has three implications for global developers:
1. Fragmentation accelerates innovation in both directions.
When the Chinese AI ecosystem operates independently, it develops its own benchmarks, evaluation standards, and best practices. Techniques that emerge from this ecosystem — like DeepSeek's Mixture-of-Experts routing or Qwen's long-context optimization — will increasingly differ from Western approaches. Developers who track only Hugging Face will miss innovations from ModelScope.
2. Maintenance burden increases for cross-border teams.
Teams that need to serve both markets face a growing operational tax. Models hosted on Hugging Face may not be accessible from Chinese infrastructure. CI/CD pipelines that rely on GitHub Actions may not reach Chinese deployment targets. Package registries split. Every tool in the stack now has two versions.
3. The concept of 'open source' diverges.
The Chinese open-source ecosystem operates under different licensing norms, different expectations around commercial use, and different regulatory requirements. What 'open source' means on ModelScope is not identical to what it means on Hugging Face. Companies building on either ecosystem need to understand both sets of rules.
This is not a prediction of permanent separation. Regulatory and market conditions could shift. But for now, the technical reality is that two parallel open-source AI ecosystems are developing, each with its own infrastructure, its own community norms, and its own trajectory. Developers who understand both will have a significant advantage.
📊 Know your real AI costs
Our TCO calculator for AI model migration covers context overage modeling, data transfer estimation, and total cost forecasting for teams evaluating model switches.
Get the framework →
This analysis focuses primarily on the open-source AI ecosystem between the U.S. and China. The assessment of Hugging Face accessibility and GitHub block patterns is based on publicly available reports and may not reflect real-time conditions given the rapidly changing regulatory landscape. The characterizations of DeepSeek, Qwen, and other Chinese model ecosystems are drawn from company-published materials and Chinese media reports, which may carry inherent biases. The "parallel ecosystem" thesis is an analytical framework, not a prediction of permanent divergence — regulatory and market conditions could shift in either direction.
Disclaimer:
The analysis above is based on publicly available data as of 2026-07-03. All benchmark scores, pricing, and performance claims are sourced from the respective companies' published materials. I am not affiliated with any of the companies mentioned unless explicitly stated. For the most current information, please visit the official sources linked throughout this article.