China weighs blocking downloads of its top AI models, not APIs
Beijing's MOFCOM is consulting Alibaba, ByteDance and Zhipu; already-downloaded weights stay yours, but future self-hosting and fine-tuning could vanish.
Copy markdown
Hosted stays open, downloads get the leash
MOFCOM is reportedly consulting Alibaba, ByteDance and Zhipu on a tiered export regime: overseas users keep API access to Chinese models while downloading raw weights for local deployment gets restricted. Nothing is law yet — it's under industry review with no timetable.
Cache the weights you depend on now
If your stack self-hosts or fine-tunes GLM, Qwen, Kimi or DeepSeek, the exposure is future frontier drops going API-only. Weights already pulled under MIT or Apache licenses stay usable, but new local deployment and derivative work on controlled models could stop.
Chip designs are on the same list
The proposed package would also bar TSMC and Qualcomm from fabricating chips designed by Chinese firms and tighten review of foreign acquisitions of Chinese AI startups — Washington-style export controls pointed outward.
Why now: China owns the open-weight lead
This lands as Chinese labs top open-weight coding — LongCat, GLM-5.2 and Kimi have led leaderboards this month. Turning downloads into a controlled export converts that lead into leverage and nudges builders toward hosted APIs or whatever weights they've already cached.