Open-Weight Models Aren’t Enough. We Need Truly Open Source AI Models for Science and Society.
China’s AI model capabilities are rapidly catching up with those of the United States. New open-source weight models such as Moonshot AI’s Kimi K3 and Alibaba’s Qwen3.8-Max have parameters reaching 240 trillion, further narrowing the performance gap. James Landae, chairman of Stanford HAI, pointed out that although these models can be downloaded and run, opening only the weights without training data, code, and tools means that true “openness” cannot be achieved, creating a gap in trust and auditing. U.S. policy is shifting from encouraging innovation to strengthening security reviews, even forcing some models to operate offline, and considering extending such restrictions to Chinese models. In response, more than 20 companies including Nvidia and Microsoft issued a public letter opposing premature restrictions, emphasizing the need for global market competition. Landae advocated for establishing a “open-source science” level based on Linux Foundation standards, with full openness in code, data, and tools, to support external research, auditing, and cultural hypothesis analysis.