The Chinese government has spent three years giving away its most capable artificial intelligence to anyone to use. That generosity was always temporary. Chinese authorities have reportedly been meeting with Alibaba, ByteDance and the start-up Z.ai to discuss restricting overseas access to some of China’s most advanced models — including systems that have not yet been released.
The U.S. government’s instinct will likely be to answer with a ban of its own. That would be a mistake. The right response is the opposite: America and its allies should build and deploy their own so-called open-weight models, which users can download, run and modify on their machines. They should do so at scale, and commit to releasing new and better versions on a dependable schedule — so that anyone building on an American model can count on the next one arriving. China’s move is an opening for the United States to take the lead again on A.I. It will not stay open for long.
An A.I. model’s “weights” are the billions of individual settings acquired and fine-tuned during training that help determine how a model responds to prompts. When a company keeps its weights secret, as OpenAI does with GPT, other companies and developers can rent access through its servers — but they cannot look inside or take the model with them. When a company publishes these weights, as the Chinese company DeepSeek has for its A.I. model, any user can download and run the weights on its own hardware, inspect them and retrain them for a specialized task.
That is how China is arguably winning the A.I. race. Last week, the Chinese start-up Moonshot A.I. unveiled Kimi K3, which the company says is the world’s largest open-weight A.I. model and rivals the performance of Anthropic and OpenAI’s best models to date. Alibaba’s Qwen is now the most downloaded A.I. model family on earth. Developers have used it to build more than 100,000 specialized models of their own. On OpenRouter, a service that lets developers send their work to whichever A.I. model they choose, Chinese open-weight models accounted for more than 20 percent of the units of text processed over a number of weeks in 2025 and more than 40 percent over a few weeks this spring; in 2024, such models’ weekly processing of text was as low as about 1 percent.
Much of that traffic is American. Airbnb’s customer service agent runs substantially on Qwen; Airbnb’s chief executive, Brian Chesky, has called the model fast and cheap. Cursor, among the most valuable A.I. coding start-ups in the country, built its Composer 2 model on a base licensed from Moonshot. DoorDash now routes the “lower-level work” of its code review system to Kimi, DoorDash’s co-founder Andy Fang reports, reserving an American model for the hardest tasks. More American businesses, attracted by lower costs, are joining these start-ups in using these models.
This had been China’s plan since 2017: push Chinese models into every market, shape the standards others adopt and pull the world toward the Chinese hardware the models run best on. The country ran the same sequence with solar and telecom equipment and, most notably, critical minerals — building the world’s dependence on cheap rare earths before restricting exports once that dependence was real.
