China brings open weight AI models under security filing, Alibaba Qwen and DeepSeek add download gates
Kai Tanaka ·
China has moved open weight artificial intelligence models into its existing generative AI security filing regime, a shift published late Tuesday in Beijing and effective immediately with a transition window into early 2027. Within hours, Alibaba Cloud’s Qwen team and startup DeepSeek adjusted download flows on domestic hubs to add identity verification and compliance attestations, according to notices on ModelScope and project repositories.
The change matters for developers and vendors outside China because many widely used Chinese models are distributed in open form and mirrored on global platforms. Gating and new license riders will affect how research teams and startups integrate Qwen and DeepSeek weights, and may limit redistribution via sites such as Hugging Face from mainland contributors.
What changed The Cyberspace Administration of China, with the Ministry of Industry and Information Technology and other agencies, issued a notice clarifying that publishing or mirroring model weights counts as providing a generative AI service if the models can be readily deployed by end users. Providers must complete security filings before public release, implement real name checks for downloads, log distribution, and include model cards with safety commitments and takedown contacts.
The document states that domestic platforms hosting weights must screen model outputs in hosted demos and provide mechanisms to suspend downloads if models are used to generate prohibited content. The notice places responsibility on mirrors, which brings Chinese developer platforms under the same enforcement expectations as primary model publishers.
Immediate platform moves ModelScope, Alibaba Cloud’s model hub, said Wednesday it will require verified accounts and usage pledges for select large language and vision weights, and it will restrict anonymous downloads for mainland IP addresses. The Qwen team told users it is consolidating distribution on ModelScope for compliance and will remove legacy third party links in coming days.
DeepSeek updated its repository to add a China specific compliance clause that bars redistribution that bypasses platform level checks. Baidu’s AI Studio and ByteDance’s Volcano Engine limited some model weights to enterprise verified tenants, and indicated that filings would be referenced on project pages once approved.
Why this matters globally Open releases from China have impacted worldwide AI development, with Qwen variants and DeepSeek models ranking high in open benchmarks and adoption charts. Stricter gating and mirror liability introduce friction for cross border collaboration, raise compliance costs for integrators, and could push some Chinese teams to release only through domestic channels that segment access by jurisdiction. For global AI vendors and cloud platforms, this signals further fragmentation of the open model ecosystem along regulatory lines.