Washington weighs banning Chinese open weights it can't recall
The Kimi K3 drop reportedly revived a US push to restrict Chinese models; 25 firms signed a letter against it, and OpenAI, Anthropic and Google didn't.
Copy markdown
The push, and what it could hit
After Kimi K3 hit frontier-level coding, the Trump administration reportedly revived a plan to restrict Chinese open-weight models — Kimi, DeepSeek and Qwen among them — with Commerce said to be weighing adding Chinese AI labs to the Entity List, which could require a license to use them.
Weights don't have an off switch
A chip is a chokepoint and a closed model is an API gate; an open weight is neither. Once Kimi K3 or DeepSeek weights spread across Hugging Face, torrents and laptops within 48–72 hours, a domestic ban becomes nearly impossible to enforce.
25 firms signed a letter; three didn't
Nvidia, Microsoft, Meta, IBM, Dell, Palantir and Hugging Face signed a July 24 letter urging against “premature restrictions” on downloadable models and against treating distillation as theft. OpenAI, Anthropic and Google are absent; Nvidia says the coalition has since grown toward 50.
Where a ban would actually bite you
Analysts peg the enforceable target as government and Fortune 500 procurement, not solo devs. But with US closed models already gated by a June executive order, curbing Chinese open ones could push builders toward open weights, not away. If you ship on Kimi or DeepSeek at a regulated employer, watch procurement policy.
The security case, contested
Treasury's Scott Bessent cited US LLM watermarks found in Chinese systems as evidence of IP theft; Nvidia's Jensen Huang called backdoor claims “misconceptions.” The gap driving the urgency: Nvidia's Nemotron 3 Ultra (550B) scored 47.7 on intelligence versus Kimi K2.6's 53.9.