CHINA WATCHOpen Models Are Becoming a Tool of Chinese Statecraft
The argument in Washington over the past few months has been about whether to ban Chinese AI models or restrict them. What has been less examined is what Beijing intends to do with the open-weight – that is, publicly available and downloadable – models its labs have built.
The argument in Washington over the past few months has been about whether to ban Chinese AI models or restrict them. What has been less examined is what Beijing intends to do with the open-weight – that is, publicly available and downloadable – models its labs have built. Beijing has started to answer that question, and not with one voice.
Xi Jinping has encouraged open source without conditions. A new intergovernmental organization was founded in Shanghai to carry Chinese AI abroad. The planning ministry tied open-source sharing to compliance systems. And the standards body began drafting rules for how open models are released and reused.
On 16 July, the UN secretary-general attended as 29 countries in Shanghai signed the founding agreement of the World Artificial Intelligence Cooperation Organization (WAICO), which will be headquartered in the city.
What China brought to the Shanghai meeting seemed less like a plan than a possession: the largest collection of open-weight AI models in the world. These are models that are already lodged in the infrastructure of countries that compete with China. What July gave this collection was a political sponsor and an institution to carry it abroad. An institution that can spread a resource, though, can also meter it.
These open-weight models are driven by two tendencies that deserve more attention in tech policy circles. The first is pace. Chinese AI labs DeepSeek, Alibaba, Zhipu, MiniMax and Moonshot ship capable models into the open every few months, each generation folding in whatever the last one learned, borrowed or invented.
The second tendency runs in a direction not many people expected. The weights enter advanced democracies. Countries without indigenous frontier-model labs must build on somebody else’s foundation, and increasingly these foundations are Chinese open-weight models. Rakuten AI 3.0, launched in March as Japan’s largest model and funded in part by Geniac, the government’s national AI program, turned out, on inspection of its configuration files, to be built on DeepSeek V3. Another Geniac model, created by the lab Abeja and also funded by the Japanese government, is built on Alibaba’s Qwen. Singapore’s flagship model, Sea-Lion, is also built on Qwen.
Seen from Beijing, the logic is not hard to follow. Under US export controls and company-blacklisting measures designed to slow China down, the open-weight route has worked. It generates domestic innovation, it allows Chinese models to travel abroad without an export license and it gives China a lever on AI
