China's draft AI export controls draw a moving frontier, not a wall

One export-control text now draws a line that frontier AI crosses but open weight does not. The joint commerce-and-science ministry notice from Beijing (MOFCOM-MOST), issued October 16, writes that line down explicitly.

The catalogue's first two thresholds are mechanical: training compute at or above 10²⁶ floating point operations, or parameters at or above 2 trillion. The third threshold is the live one, defined as "globally leading capability" by top-10 standing on a public assessment at the time of export, covering specialized capability, usage, safety, or training efficiency, not only overall rank. A model can be sub-trillion parameters and still land inside the restricted bucket if a regulator points to a single benchmark where it sits in the top ten.

The MOFCOM-MOST notice writes the carve-out directly into the rule: technologies that have entered the public domain through open-source release or published academic papers sit outside the management scope. That sentence is the textual basis keeping existing Chinese open-weight releases outside the restricted set, regardless of how the third threshold is later read.

Export controls written in compute and parameter units freeze yesterday's frontier; the third threshold turns the rule into a moving target that follows the leaderboard, because scope is set by where a model sits on a public assessment at the moment of export rather than by where it sat when the catalogue was written. Builders, labs, and trade partners have 30 days, the consultation closes November 15, to push for a narrower definition of "globally leading" before that moving target hardens.