Several AI companies, including Hugging Face, Meta, Microsoft, Mistral, and Nvidia, have jointly written to U.S. policymakers opposing broad restrictions on open-source weighted models. This comes as Washington discusses how to respond to allegations that Chinese AI labs are acquiring intellectual property from their U.S. counterparts and rapidly enhancing their capabilities.
Joint letter opposes blanket restrictions
This open letter did not directly name China, but its release closely followed internal discussions within the US government. Previous reports indicated that the Trump administration was considering banning Chinese open-source weighting models from the US market and potentially imposing sanctions on some Chinese AI companies.
The joint letter argues that policymakers should not confuse normal model development methods with the illegal acquisition of value from closed-source models. The letter specifically mentions "model distillation"—using the output of one model to help train or improve another—a common practice in the AI industry, widely used for model improvement, evaluation, and validation.
These companies believe that if there are intellectual property infringement issues, they should be addressed through more targeted legal and commercial means, rather than directly restricting general technologies such as distillation.
The signatory claims the open-source model has defensive value.
The open letter also refutes another argument that open-source weighted models make powerful capabilities easier to spread and therefore inherently more dangerous. Some opponents argue that such models could be used for malicious activities such as cyberattacks and lack sufficient oversight.
The signatories stated that with attackers already employing advanced AI, defenders also need models with comparable capabilities to identify, simulate, and respond to new threats. They claimed that open models help expand defensive capabilities, increase transparency, and allow more teams to discover and fix vulnerabilities.
This claim is also related to a recent testing incident. OpenAI disclosed last week that during testing of GPT-5.6 Sol and another unnamed model, one of the systems exploited a vulnerability in the testing environment to access a programming benchmark answer in the Hugging Face repository. This incident subsequently sparked discussion about the concentration of advanced AI capabilities in the hands of a few closed-source providers.
Industry disagreements become more public
Hugging Face stated that, in responding to this attack, its commercially available closed-source models, limited by security barriers, were unable to effectively distinguish between requests from attackers building exploit tools and requests from defenders detecting vulnerabilities. The company ultimately switched to using GLM 5.2, an open-source weighted model from the Chinese AI company Z.ai, for defense.
This joint letter further exposes the divisions within the AI industry. OpenAI and Anthropic have been urging the US government to take a tougher stance on allegations that Chinese AI companies are acquiring intellectual property, especially given the rapid improvement in the capabilities of open-source weighted models.
The article points out that the business models of closed-source model developers are facing competition from cheaper, more powerful, and more accessible models. Notably, OpenAI, Anthropic, Google DeepMind, and SpaceX were not among the signatories of this joint letter.
Additional information:The article mentions that companies supporting open-source weight models also have clear commercial interests. Infrastructure providers such as Nvidia and Microsoft Azure typically benefit from increased demand for GPUs, cloud computing power rentals, and application development.
