At a Rest of World gathering in San Francisco, experts highlighted that Chinese open-weight artificial-intelligence models provide high capability while costing far less than the United States’ frontier systems. Despite the United States and Europe deploying greater financial and computational resources, their newly released open models still trail the dominant Chinese offerings in performance, according to the speakers.
Last week American startup Reflection unveiled its first open-weight model, while European labs Aleph Alpha and Mistral also introduced new versions. Open models such as these grant users greater customization and control compared with the closed offerings of OpenAI, Anthropic and Google, and they do so at substantially lower operating costs, the panel noted.
Cost considerations dominate adoption decisions for businesses and governments in lower-income nations, according to Daniel Björkegren, an assistant professor at Columbia University. He explained that beyond price, the expertise required to adapt open models for specific sectors,such as education, healthcare and agriculture,is a critical factor. The lower expense of Chinese models, he said, makes them especially attractive for these contexts.
Chinese President Xi Jinping has framed open models as essential to the nation’s ambition of achieving global AI leadership by 2030. In line with that vision, China created the World Artificial Intelligence Cooperation Organization this year to promote AI uptake and influence governance in developing countries. The initiative includes providing access to Chinese models and delivering training for students and workers, according to event speakers.
The United States recently ordered Anthropic to block foreign nationals from its Fable 5 and Mythos models on national-security grounds, a move that reinforced concerns among several governments about over-reliance on American systems. Björkegren cited the directive as evidence that nations are looking for alternatives, and that Chinese open models could fill that strategic gap.
Following Reflection’s announcement, Nathan Lambert, head of AI research at Trillium Labs, posted on X that the startup joins Nvidia and Thinking Machines in releasing strong models yet still trails Chinese counterparts. He warned that even well-funded U.S. researchers may fail to win, emphasizing the difficulty of competing with China’s rapid progress in large-language-model development.
Chinese AI firms continue to wrestle with compute constraints and revenue generation. Their open models can perform certain tasks at 50 % to 90 % lower cost than leading American frontier models, according to the speakers. Moonshot AI is negotiating revenue-sharing agreements with Microsoft, Amazon and Google to host its Kimi K3 model, marking a rare collaboration between a Chinese AI company and major U.S. cloud providers.
Monetization remains a central concern for Chinese developers, said research fellow Kristy Loke of the MATS program. She explained that the strategy is to achieve broad adoption first, hoping that satisfied users will remain loyal and eventually generate revenue. Loke added that success with open models serves both domestic soft-power goals and international market expansion for China.