A July 24 industry letter urges U.S. policymakers to avoid premature restrictions on open-weight artificial intelligence, arguing that downloadable models can widen access and give users more control over deployment.
Microsoft, which hosts the letter, said more than 270 organizations had signed by Aug. 3. The signatories include NVIDIA, Microsoft, Meta and Google.
The coalition says organizations can tailor downloadable models to their needs, avoiding the expense of training from scratch and frontier-model prices for every task. It argues that startups, businesses, universities and public institutions could benefit from wider access, while users gain more control over their data and where models run. The letter draws an analogy to the open-source software movement. These are the signatories’ arguments for open weights, not reports of realized savings.
The letter also describes a trade-off: once weights are released, developers may have limited oversight of how they are reused, and later versions can be difficult to identify. That leaves model creators with less control as others adapt their systems.
Recent releases from Chinese developers show the range of models and terms involved. Moonshot AI describes Kimi K3 as a 2.8-trillion-parameter model, with 104 billion parameters activated and a context window of 1,048,576 tokens. Its full weights are available under Moonshot’s Kimi K3 License. Moonshot acknowledges that Kimi K3 has not matched top proprietary models overall.
Alibaba’s Qwen team lists Qwen3.8-27B at 27 billion parameters, with a native context length of 262K tokens that can be extended to 1 million using YaRN. The team says its weights became available on Hugging Face and ModelScope on Aug. 14 under the Apache 2.0 license.
The policy debate also includes U.S. security concerns. In an April 23 memorandum, the White House alleged that foreign entities, principally based in China, were conducting industrial-scale distillation of U.S. frontier AI systems. The memorandum also described legitimate distillation as important to the AI ecosystem.
A June 2 executive order called for agencies to develop a voluntary evaluation process for covered frontier models. It describes a possibility in which developers give the federal government access as much as 30 days before a planned release to other trusted partners, and separately discusses choosing trusted partners for early access. The order says that section does not authorize mandatory licensing or preclearance for model releases.
