Open-weight AI models face testing exemption and new licensing conditions
The White House signaled open-weight models will skip voluntary safety tests, while MiniMax restricted free local self-hosting for H3 in four markets and Mistral AI released its Shieldstral open-weights safety classifier.
Artificial Intelligence··Midday
Open model exemption in voluntary tests
The Trump administration informed artificial intelligence developers that open-weight models will not undergo the White House voluntary cyber safety tests applied to closed systems. According to Reuters, White House officials discussed unpublished testing rules with staff from Meta, Anthropic, Google, Nvidia, and OpenAI. The meetings followed disclosures that tools from OpenAI and Anthropic had breached other companies' systems during evaluations. The administration had stated in June that the tests would remain voluntary and focus primarily on models with sophisticated hacking capabilities. Officials cited Nvidia's Nemotron and Meta's Llama as examples of open models, while categorizing systems from OpenAI, Google, and Anthropic as major US closed-model developers. Sharing the testing framework with a small group of companies without public disclosure drew criticism; Americans for Responsible Innovation argued that keeping criteria confidential deepens a gap in federal oversight. While the White House media office did not respond to requests for comment, 5 Democratic senators called for legislation making testing permanent for the most advanced American-made frontier models.[1]
Carve-out licensing regime for MiniMax H3
Shanghai-based artificial intelligence startup MiniMax released weights for its H3 video model to developers while imposing a regional licensing requirement for free local self-hosting. According to the South China Morning Post, the released weights allow developers to use, copy, modify, and distribute the software without paying MiniMax; however, the company restricted free local access in the United States, the European Union, the United Kingdom, and South Korea. Developers in these 4 markets must submit a formal license application to self-host the software on local infrastructure without paying fees. MiniMax said it may authorize use after reviewing deployment scenarios and confirming appropriate compliance controls and safeguards. The company specified that paid API access via its website or authorized third parties remains available worldwide. MiniMax attributed the regional carve-out to regulatory uncertainties and ongoing copyright-related legal proceedings. The release highlights that open weights do not automatically guarantee unrestricted free local deployment, as providers pair model weight releases with geographic licensing approvals.[2]
Mistral releases open inference safety classifier
France-based Mistral AI released Shieldstral, a 3 billion parameter open-weights multimodal safety classifier published under the Apache 2.0 open-source license. According to the company, the model treats each request as an instruction, a yes or no question, and a document to judge—a prompt, a response, a pair, or an image—evaluating plain-language policy questions at inference time without retraining. Returning calibrated safety scores for text and images directly from yes and no logits, the model avoids a fixed harm taxonomy baked into the weights, allowing deployment policies to be rewritten dynamically at inference. Mistral claims Shieldstral runs on a single 16 GB NVIDIA GPU and matches text safety performance of open guard models up to 7 times larger. The release was framed alongside Mistral's membership in the Open Secure AI Alliance with Nvidia and other organizations. As Washington signals open-weight models will skip voluntary cyber tests and MiniMax subjects free local hosting to licensing conditions, open safety classifiers like Shieldstral enable developers to implement custom moderation controls within local deployment environments.[3], [1], [2]
Related columns
For more information on this topic, you can read the related columns.
