How China overtook the US in the global market for open AI models

.

China has pulled ahead of the United States in one of the most important and fastest-growing corners of artificial intelligence: open models. According to new research by the Massachusetts Institute of Technology and open-source platform Hugging Face, Chinese-made open AI models accounted for 17 per cent of global downloads in the past year. That figure surpasses the 15.8 per cent share of downloads from American developers such as Google, Meta and OpenAI. It is the first time China has taken the lead in this segment, marking a significant shift in global AI influence, the Financial Times reported.

Open models — whose weights can be downloaded and modified freely — form the backbone for thousands of downstream applications. Start-ups use them to build products quickly and cheaply, researchers use them to test new ideas and governments use them to develop sovereign AI systems. Whichever country dominates this layer of AI infrastructure gains substantial global influence.

The study highlights a stark divergence in strategy. China has embraced the release of open models, encouraged directly and indirectly by Beijing. The US, by contrast, has become more cautious, with its biggest companies pushing deeply into closed, proprietary systems that maximise commercial control.

OpenAI, Google and Anthropic have built powerful frontier models but restricted access through APIs, subscription platforms or enterprise contracts. Their most advanced systems—including those under development in 2025—remain entirely closed, with neither weights nor training data made public.

China’s major AI groups have taken a very different approach. DeepSeek and Alibaba Cloud, the two most downloaded contributors among Chinese developers, release frequent iterations of open models that are smaller, faster and easier to fine-tune. These models have become favourites among smaller companies and independent developers across Asia, Africa and Latin America.

DeepSeek’s R1 model, unveiled earlier this year, jolted the global AI community. Despite being trained at a fraction of the cost and computational power of US rivals, R1 demonstrated sophisticated reasoning capabilities that rivalled much larger American models. Its release triggered widespread discussion over whether US labs, despite their vast compute budgets, could maintain their competitive edge.

Alibaba’s Qwen series has also become a dominant force in the open-model ecosystem, offering variations optimised for text generation, coding tasks and multimodal work. Both companies have adopted a rapid release schedule, shipping new models weekly or biweekly, in stark contrast to the six-month or yearly cycles typical among large US labs.

US export controls on advanced Nvidia chips have limited China’s access to top-tier hardware, but analysts say this has driven innovation rather than hindered it. Chinese researchers have

been forced to optimise, relying on techniques such as model distillation to shrink large systems without major performance losses.

This resource-driven creativity has allowed Chinese developers to release highly efficient models that perform strongly on consumer hardware, widening their global adoption. Several leading Chinese labs have also moved aggressively into video generation, an area expected to dominate AI investment in the coming years.

China’s dominance in open models carries geopolitical implications. Researchers studying these systems have identified clear political biases aligned with the Chinese Communist Party. Chinese models tend to avoid generating sensitive content related to Xinjiang, Tiananmen Square or Taiwan, and their factual outputs often mirror official narratives. As these models spread globally, they shape information flows far beyond China’s borders.

Meanwhile, American companies are pouring resources into frontier AI projects, racing to build systems capable of what they describe as “superintelligence.” Meta, once the leading champion of open-weight models through its Llama series, has begun shifting toward more closed systems as it competes with OpenAI and Anthropic. OpenAI has taken small steps toward openness with limited “open weight” releases, but these fall far short of full transparency.

Only a handful of American organisations — most notably the Allen Institute for AI, which released the fully open-source Olmo 3 model — continue to champion openness.

Analysts warn that China’s rapid ascent in the open-model landscape should concern US policymakers. If open systems become the default infrastructure for AI innovation worldwide, the country that shapes them will have disproportionate influence over the future of software, research and global information ecosystems.

The US remains a leader in frontier AI research, but China now dominates the layer of the industry where global adoption happens at scale. The question facing Washington is whether it can—or will—pivot before China consolidates its lead.