Chinese artificial intelligence models have overtaken American developers in global downloads for the first time, capturing 17 percent of the market compared to 15.8 percent for U.S. creators, according to a joint study by the Massachusetts Institute of Technology and Hugging Face. The shift marks a pivotal moment in the race to shape AI’s future, driven largely by DeepSeek and Alibaba’s Qwen models, which have gained traction by offering cost-effective alternatives with frequent updates.
Security Vulnerabilities Raise Alarm#
The surge in Chinese model adoption has sparked serious security concerns. Research published November 20 by cybersecurity firm CrowdStrike revealed that DeepSeek’s AI models produce significantly more insecure code when prompted with politically sensitive references to Tibet, Uyghurs, or other topics deemed problematic by Beijing.
While DeepSeek generated vulnerable code in 19 percent of neutral cases, that figure jumped to 27.2 percent when asked to write code for industrial control systems in Tibet. The flaws appear to emerge after the model’s reasoning process completes, suggesting embedded biases rather than poor training data.
Studies have also documented clear ideological alignment with the Chinese Communist Party, with models refusing to produce content on sensitive subjects like Taiwan or the Tiananmen Square incident. “It should raise alarms in the U.S. that China is making substantial progress in the open model domain,” said Janet Egan, a fellow at the Center for New American Security.
Strategic Divergence Fuels Competition#
China’s open-source strategy contrasts sharply with leading U.S. tech companies. OpenAI, Google, and Anthropic have maintained tight control over their most advanced models, focusing on developing artificial general intelligence through closed systems and generating revenue via subscriptions. Meta, previously a champion of open-source AI with its Llama models, has shifted toward developing closed models as it pursues “superintelligence.”
Chinese firms have been pushed toward open-source releases partly due to U.S. export restrictions on advanced Nvidia chips. This has forced Chinese developers to adopt more innovative approaches, releasing model variations weekly or biweekly compared to the semi-annual or annual cadence typical of U.S. labs.
The practical impact is already visible: up to 80 percent of AI startups pitching to venture capital firm Andreessen Horowitz now use Chinese open-source models, according to partner Martin Casado. The Allen Institute for AI released Olmo 3 in November as one of the few major American contributions to open-source AI, underscoring the growing gap in this critical sector.