- Lead. China’s Ministry of Commerce has held talks with Alibaba, ByteDance and Z.ai about restricting foreign downloads of AI model weights — a move that would reverse Beijing’s strategy of offering cheap open-source alternatives to American systems.
- Fact. The discussions, which took place over roughly the past month, covered a tiered regime: simple filing for basic tools, security reviews for stronger models, and a possible domestic-only lockdown for the most capable systems.
- Stake. Developers across Europe and Asia currently rely on Chinese open-weight models such as Alibaba’s Qwen and ByteDance’s Doubao as cost-competitive substitutes for US frontier AI; curbing access would reshape the global model landscape at a moment when both Washington and Beijing are tightening AI controls.
Chinese authorities have been consulting major domestic AI developers — including Alibaba, ByteDance and Z.ai — about the possibility of restricting overseas access to China’s most advanced artificial intelligence models, according to The Next Web, citing sources familiar with the discussions. The Ministry of Commerce led the talks, with the planning agency NDRC also involved. No decisions have been finalised, and any curbs under discussion might apply only to future model releases.
What Is Under Discussion
The proposed framework covers two distinct concerns. The first is the transfer of training data abroad — a question Beijing has been pressing since at least 2023 under its broader data-governance rules. The second is whether foreign users should continue to be able to freely download the weights of Chinese AI systems — the parameters that make a model run locally without an API.
A tiered classification was floated: the least capable open-source tools would face simple registration requirements; intermediate systems would require security reviews before release; and the most capable models could be restricted to domestic deployment only. API and cloud-based access, where usage can be monitored, would remain available under the scenarios discussed.
The Models at Stake
Alibaba’s Qwen series, ByteDance’s Doubao, and Z.ai’s GLM-5.2 — whose coding benchmark performance surpassed GPT-5.5 at a fraction of the cost — sit at the centre of the consultation. All three circulate globally as open-weight downloads, making them accessible to any developer willing to run their own inference rather than paying for US cloud APIs. DeepSeek’s models are similarly positioned, though the reports do not specify whether DeepSeek participated in the talks.
The potential impact extends beyond Chinese companies. European AI developers and startups in South-East Asia have built products on Chinese open-weight models precisely because they are free to download and can be fine-tuned without the usage restrictions attached to commercial US APIs. A weight-download ban would force those developers to either pay for cloud API access or shift to Western models at higher cost.
The Geopolitical Signal
The consultations align with China’s parallel push into multilateral AI governance. The 29-nation AI cooperation body Beijing launched at the Shanghai conference positions China as a rule-setter in the Global South — a role that depends on Chinese models being widely accessible and trusted. Restricting model weights would complicate that narrative while simultaneously treating advanced AI as a strategic national asset to be guarded, not shared. The tension between those two postures has not been resolved, which may explain why no implementation timeline has emerged.
If Beijing does move forward, it would create a near-mirror image of the US export-control framework: two of the world’s largest AI producers each restricting the other’s developers from accessing their most capable systems, fragmenting a model ecosystem that has so far remained more open than the chip supply chains beneath it.