Chinese AI leaders debate open-source risks and benefits at Wuzhen Summit

The drive for faster AI

Researchers from three major Chinese AI developers gathered at the 2026 World Internet Conference Wuzhen Summit to discuss the future of open-weight models. The session, titled the Wuzhen Dialogue on Artificial Intelligence, featured leaders from DeepSeek, Alibaba's Qwen team, and MiniMax. This year's summit marked a significant shift by making open source a primary theme of the official programming.

DeepSeek research engineer Xi Gao explained that the primary motivation for releasing models openly is the pursuit of artificial general intelligence. He noted that the most significant competitor for AI laboratories is time rather than other specific companies. By opening their research, labs allow capable peers to build upon their work, which can accelerate the overall pace of development.

While technical barriers can provide a temporary advantage for a single company, Gao suggested that these hurdles are eventually overcome. He argued that inviting the broader community to participate in building better models is a more effective way to move the needle. This approach prioritizes collective progress over maintaining a closed, proprietary advantage.

Scaling the open ecosystem

Alibaba's Qwen team provided a look at the massive scale of the current open-source landscape. Dayiheng Liu, the Qwen LLM project lead, shared that the team has released more than 460 models since its first generation launched three years ago. These models have garnered over 4 billion downloads globally and have served as the foundation for more than 330,000 derivative models.

Liu emphasized that the open-source model is designed to make AI more accessible to everyone. By allowing the public to participate in building AI, the ecosystem fosters a wider variety of applications and innovations. This democratization allows developers to use the models while also contributing to the underlying technology that powers them.

MiniMax chief architect Wen Feng added that their strategy involves providing differentiated value through multi-modal models. He noted that the company shares not just the final products, but also the engineering experience behind them. This includes sharing data, tool modifications, and the specific problems encountered during the experimental process.

The challenge of irrevocability

Despite the benefits of open access, the panel addressed the specific risks that come with open-weight models. Liu identified a major weakness as irrevocability, which refers to the fact that once a model is released, it spreads widely. Unlike hosted models, which can be patched or updated by the provider, open models cannot be easily retracted or modified once they are in the hands of the public.

Another significant concern is the lack of external guardrails for these models. Because safety depends on built-in safeguards, third-party developers may attempt to bypass the restrictions set by the original creators. Liu argued that model restrictions alone are insufficient for governance and that a combination of laws, industry norms, and supporting infrastructure is required.

However, the panel also suggested that openness can be a tool for safety. Because these are white-box models—meaning their internal structures are visible—researchers can study them more thoroughly to identify flaws. Furthermore, open models can be deployed specifically for cyber defense purposes to protect against digital threats.

Industry needs and safety

For many companies, open models are often the preferred choice because they offer a high degree of control. Feng noted that businesses can fine-tune these models to suit specific tasks and deploy them privately within their own systems. This allows for a level of customization and data privacy that is often difficult to achieve with closed, proprietary APIs.

The discussion also touched on the potential for smaller models to be more capable than previously thought. Liu pointed out that even models with as few as 1.5B parameters can be unlocked by new post-training techniques. This suggests that the capabilities of smaller, open-source models might be underestimated by those who only focus on the largest systems.

DeepSeek acknowledged that open-source models can amplify safety risks and stated that the company feels a responsibility to prevent these risks from spreading. Gao mentioned that the team is currently researching solutions to mitigate these dangers. The goal is to balance the benefits of open collaboration with the necessity of keeping AI systems safe for public use.