China proposes export curbs on large language models and chip designs

China's Ministry of Commerce and Ministry of Science and Technology have released a draft proposal to restrict the export of 23 artificial intelligence technologies. The measures cover frontier large language models, autonomous agents, computing clusters and semiconductor design tools.

The draft forms part of a wider update to China's catalogue of controlled technologies, which also lists items in biomedicine and solar manufacturing. While the public comment period runs until 15 November 2026, an interim pre-export registration requirement for the named technologies took effect immediately on Friday, 16 October 2026.

Limits placed on weights and source code

Under the proposed rules, Chinese regulators intend to restrict overseas transfers of core software assets, including model weights and underlying source code. The restrictions would apply to models that reach 2 trillion parameters, match training compute thresholds of 10 to the power of 26 floating-point operations, or rank in the top 10 on global benchmark evaluations. Frontier systems from Chinese labs, including Alibaba's Qwen family and Moonshot AI's Kimi releases, fall into these categories.

The draft also seeks to restrict chip design blueprints for processors made at 14 nanometres or below, as well as advanced 2.5D and 3D stacking techniques. In a joint statement, Chinese officials stated the controls aim to protect national security and prevent malicious uses against critical digital infrastructure.

Importantly, the proposal draws a line between technology exports and online services. Chinese authorities noted that foreign users can still access models through cloud application programming interfaces, web portals and mobile apps without triggering export controls. Software and research already in the public domain remain exempt, and officials indicated that open-source releases may be granted fast-track approvals once registered.

What this means for Australian businesses

Many Australian developers and small enterprises turn to Chinese open-weight models, such as DeepSeek or Alibaba's Qwen, because they offer strong performance at low cost. The draft rules will not cut off existing open-source versions that teams have already downloaded or deployed on internal systems.

However, Australian businesses planning their software roadmaps should prepare for potential friction in future releases:

• Cloud access remains stable. Using Chinese models through commercial cloud APIs, Australian host providers or standard consumer applications will remain allowed under the proposed rules. Day-to-day use of chat assistants and software plugins will not face immediate disruptions.
• Self-hosted systems need fallback plans. Companies that download weights to run models on local servers or private cloud instances should keep alternatives ready. If future frontier weights face export review delays, local operations should be able to switch to open models from other regions, such as Meta's Llama or European alternatives.
• Vendor checks are sensible. When hiring software contractors or buying specialised AI tools, ask suppliers where their underlying models are sourced and whether upcoming features depend on overseas model downloads.