China Dominates Open Source AI, Attracts the Global South

China Takes the Lead in Open Source AI

Chinese laboratories have made significant strides in the field of open source artificial intelligence, surpassing the United States for the first time in global downloads. According to a study by MIT and Hugging Face, reported by MIT Technology Review, Chinese open source models accounted for 17.1% of downloads on Hugging Face over the past year, compared to 15.86% for American models.

The turning point occurred in January 2025 with the release of DeepSeek's reasoning model R1 under the MIT license. This model, freely downloadable and modifiable, offers performance comparable to that of closed American models but at a significantly lower inference cost. Since then, several Chinese players such as Z.ai (formerly Zhipu), Moonshot with Kimi K2.5, and Alibaba with Qwen and MiniMax have bolstered this momentum. Notably, the Qwen family surpassed Meta's Llama in terms of cumulative downloads by the end of 2025.

According to data from a16z and OpenRouter, Chinese models account for nearly 30% of global AI usage. This gap between downloads and usage indicates that these models are not merely being tested but are indeed being put into production.

Massive Adoption in the Global South

The adoption of Chinese models is not limited to a single continent. In Africa, Southeast Asia, and Latin America, these models fill a gap left by American restrictions. American proprietary models often require credit cards, subscriptions, and network access to servers located in the United States, which is not always feasible for these regions. In contrast, DeepSeek's model is free, runs on modest hardware, and does not impose data localization requirements.

The CEO of Hugging Face also highlighted the rise of these models in Europe, where businesses and administrations are seeking alternatives to American APIs. Mistral AI, a French unicorn, is positioning itself in this niche with a "sovereign" approach. The company raised €1.7 billion by the end of 2025, built a data center in Sweden, and aims for €1 billion in revenue by 2026. According to industry estimates, 72% of large European companies deploying an open source model would choose Mistral Large.

Two Diverging Approaches

However, Mistral AI finds itself torn between open source and proprietary models. Its most ambitious business model, Mistral Large, is distributed via Microsoft Azure. On Hugging Face, downloads of Codestral (380,000 in the first week of April) lag far behind Qwen3-72B (640,000). The Chinese strategy focuses on volume and a captive ecosystem, while Mistral's approach targets compliance and the European enterprise. Two visions of open source, two definitions of sovereignty.