White House frontier AI review exempts open weights
The voluntary framework — reviewed Aug 4 with OpenAI, Anthropic and Meta but never published — asks closed labs for a 30-day pre-release hold.
Copy markdown
Open weights sit outside the whole regime
The text says nothing in it should be read as restricting open models once they're released. Build on Qwen, DeepSeek, Llama or GLM and none of the pre-release review touches you — no hold, no secure-storage mandate, no access logs.
Closed frontier models get a 30-day pre-release hold
Covered labs are asked to submit models up to 30 days before launch; during review, employee access is restricted and weights sit in secure, logged environments. The next closed GPT, Claude or Gemini could reach your API weeks after the lab finished it.
'Covered' means closed, SOTA and risky — none of it defined
The framework tags covered models as closed-source, state-of-the-art and a national-security risk, but never says what 'state-of-the-art' or 'national-security risk' mean. Which closed models actually get held is left wide open.
Voluntary, and no licensing
It explicitly rules out mandatory licensing, preclearance or permits, and taking part is voluntary. Built on June's EO 14409 with an Aug 1 deadline, it's guidance with no enforcement behind it.
And you're not allowed to read it
The White House walked through the framework Aug 4 with Meta, Nvidia, Microsoft, OpenAI and Anthropic, then declined to publish it. CFR's Chris McGuire called it "baffling": "We can't have secret, voluntary rules to regulate the most important tech in the world."