Geopolitics and the LLM open-source ecosystem: Why China blocks Hugging Face while domestic open-weights flourish

Geopolitics meets the open-source LLM crowd: Hugging Face is blocked in mainland China, even as domestic open-weights flourish. The block is network-level, not a publishing ban, and it rides on China’s AI rules about what models can generate ^[1].

Inside China, domestic ecosystems pick up the slack:

• modelscope.cn provides an accessible, China-only hub for open models, partially filling the gap left by blocks on external sites ^[1].

• GitHub remains restricted; censorship spans API/inference, and many weights are still downloadable to run locally with llama.cpp, including models like DeepSeek—though jailbreaks and filters complicate results ^[1].

Meanwhile, outside the gate, open-weight innovation keeps marching. LongCat-Flash-Thinking from Meituan is pitched as a standout among open-source models for logic, math, coding, and agent tasks, with performance and efficiency wins highlighted in circulation ^[2]. The model claims include 64.5% fewer tokens to reach top-tier accuracy on AIME25 with native tool use and Async RL delivering a 3x speedup over Sync frameworks ^[2].

Regulatory regimes will continue shaping access, development, and the global LLM landscape, with China fostering domestic ecosystems even as external access tightens.

References