"open-source and openness" — President Xi Jinping used those words at the World AI Conference in Shanghai, and the phrase has been read as a signal about the future of Chinese AI. But a careful reading of his remarks, and of how Chinese labs have behaved, shows the slogan is broad enough to accommodate both generous sharing and selective secrecy.
President Xi Jinping and the World AI Conference
At the World AI Conference in Shanghai, President Xi Jinping invoked the slogan "open-source and openness" (开源开放). Reporting and commentary quickly interpreted this as an endorsement of open-source AI models. Yet the source material makes clear that Xi did not explicitly single out model weights or frontier systems for mandatory release. A statement from the chair of the High-Level Meeting on Global Governance of AI that accompanied the conference urged an open-source ecosystem as “an important path to the inclusive development of AI,” but added that governments should “respect enterprise’s independent choices” on issues such as intellectual property.
Chinese labs are choosing both openness and proprietary release
Chinese AI developers are already demonstrating mixed behavior. Start-ups Moonshot AI and Z.ai released large open-source models this month, with Z.ai’s release arriving a day after Anthropic restricted international access to its model Fable 5. By contrast, some labs have begun keeping the weights of their latest models proprietary: ChinaTalk reported in April that a number of Chinese labs had stopped publishing weights, arguing they could not afford to train expensive models and then give them away.
Major firms follow hybrid strategies. ByteDance open-sources some models while keeping flagship image and video-generation models closed-source. Alibaba previously released weights for earlier models but kept its latest Qwen-3.7 proprietary; on July 19 Alibaba said the next iteration, Qwen-3.8, would be open-source—an apparent volte-face the source links to Xi’s call for openness, albeit noting Xi did not explicitly mean open-weights.
Closed-door talks and the Reuters 'silicon curtain' report
Concerns about restricting access have also surfaced at the policy level. Reuters reported that the Chinese government has held closed-door meetings with AI labs about restricting international access to frontier models, and used the phrase that a “silicon curtain” could rise around Chinese AI. Separately, the Reuters coverage framed Chinese leadership as “now treating cutting-edge artificial intelligence as a critical national asset that needs controls.”
The phrase 开源开放, "open-weights," and what gets shared
The Chinese term Xi used — 开源开放 — has an institutional history. It first appeared in 2017 alongside China’s initial official AI policy, and the 14th Five-Year Plan in 2021 encouraged institutions to open-source code to accelerate development. In practice, the slogan has been applied broadly by state media and officials to cloud computing, datasets, chip architecture and other AI inputs. Industry commentators note that Chinese usage does not draw the finer distinction English speakers now make between “open-source” and “open-weights.”
The result: “open-source” in Chinese official and press usage can mean many different kinds of sharing — research outputs, developer tools, datasets — without guaranteeing that the weights of the most powerful models will be published. The record already shows labs treating different components differently, and the High-Level Meeting’s language backs the idea that companies’ proprietary choices should be respected if needed.
What this means for Chinese AI labs, policymakers, and international adopters
- Chinese AI labs and start-ups: Expect continued internal tension between research teams that value openness for rapid iteration and marketing or finance teams that prefer proprietary revenue models. Moonshot’s researchers reportedly favour open-source while its marketing team prefers closed-source revenue, illustrating the trade-off.
- Policymakers and regulators in China: The government appears to be balancing promotion of a broad “open-source ecosystem” with discretion to restrict access to frontier capabilities. Closed-door discussions reported by Reuters and the High‑Level Meeting’s language suggest authorities will tolerate, and perhaps encourage, both approaches depending on national-security and commercial considerations.
- International adopters and enterprises: They can expect Chinese actors to continue releasing tools, datasets and many models that support global adoption, while also encountering proprietary frontier systems whose weights are not shared. The mix makes adoption decisions a matter of weighing access against strategic limitations, as demonstrated by Alibaba’s switch between Qwen-3.7 and the pledge for Qwen-3.8.
Strategically, the pattern the source describes is coherent: openness accelerates adoption and influence, while selective closure protects high-value capabilities. As the source puts it, China is likely to continue open-sourcing many building blocks for AI systems while reserving the weights of any “frontier AI model of unprecedented power.” That dual approach—“two legs,” in the phrasing offered by Tang Jie in People’s Daily—will shape how Chinese AI is built, sold and shared going forward.
