Update: 25 tech firms defend open weights; OpenAI, Anthropic sit out
Washington is reviving a push to ban Chinese models — the open weights your stack runs on. Nvidia, Meta, Hugging Face, Mistral signed to stop it; OpenAI didn't.
Copy markdown
The ban push is aimed at your toolkit
The Trump administration is reportedly reviving a plan to restrict Chinese AI models after this month's Kimi K3 and DeepSeek V4 releases, with Treasury's Scott Bessent reportedly citing US-model watermarks found inside Chinese systems as grounds to sanction those involved. The open weights in the crosshairs are the exact ones builders now self-host.
25 signed — the closed-model labs didn't
Signatories include Nvidia, Microsoft, Meta, IBM, Dell, Palantir, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation and Mozilla. Conspicuously absent: OpenAI and Anthropic — as Tom's Hardware notes, none of the signers sells access to a closed frontier model.
What the letter actually demands
It urges Washington to avoid “premature restrictions” on downloadable models, expand compute access for startups and researchers, and fund shared training datasets and evaluation frameworks. It also defends model distillation as legitimate, drawing a line against “unlawful efforts to extract value from closed models.”
Huang: 1 in 4 tokens is already open
Jensen Huang promoted the letter on X, arguing “the world needs both frontier closed models and frontier open models” and that “one in every four tokens generated today comes from an open model” — his frame casts open weights as this decade's open-source-software moment.
Hedge your stack this weekend
If any pipeline depends on Chinese open weights, mirror the checkpoints you rely on now — pull the Hugging Face snapshots and keep local GGUF copies — so a sudden import or hosting restriction can't strand you mid-build. Self-hosting keeps you running regardless of how the access rules land.