DeepSeek, Qwen and MiniMax Discuss Open-Source Model Risks at Wuzhen Summit

At the 2026 World Internet Conference Wuzhen Summit, DeepSeek’s Xi Gao, Qwen’s Dayiheng Liu, MiniMax’s Wen Feng and AI governance expert John Higgins discussed open-source model releases and their governance in a session titled the Wuzhen Dialogue on Artificial Intelligence link.

Gao said DeepSeek open-sources to accelerate progress toward artificial general intelligence, arguing that time, not competition, is the primary barrier. Liu noted Qwen has open-sourced more than 460 models since its first generation three years ago, achieving over 4 billion global downloads and spawning more than 330,000 derivative models. Feng explained MiniMax’s aim is to provide differentiated value by releasing multimodal models and sharing engineering artifacts, including data and tool modifications.

On risk, Liu identified two specific weaknesses: irrevocability—once a model is released it cannot be patched like a hosted service—and the absence of external guardrails, leaving safety to built-in mechanisms that third-party developers may bypass. He called for laws, industry norms and supporting infrastructure, while noting that open models also enable safety research and cyber-defense applications. Feng added that open models are often more controllable for enterprises because they can be fine-tuned and privately deployed. Gao acknowledged the amplified safety risks and said DeepSeek is researching ways to prevent harmful diffusion.