On the 4th, the White House rolled out a framework targeting leading domestic AI companies, aiming to assess the capabilities of their cutting-edge models. Notably, U.S. open-weight models were left out of this review process. Multiple sources with knowledge of the matter shared this information with U.S. media outlets on the same day.

Open-weight models, often regarded as a subset of open-source models, make their weight files—that is, the trained model parameters—publicly accessible. This allows users to download, deploy locally, and fine-tune these models without needing to disclose their training data, complete code, or the intricacies of the training process. According to these sources, only closed, proprietary U.S. models that demonstrate 'state-of-the-art' performance in cybersecurity and attack capabilities during benchmark tests are mandated to undergo voluntary government testing by their developers before being released to the public.