オープンウェイトモデルに関するAnthropicの立場Our position on open-weights models

Anthropic chief executive Dario Amodei has published a post setting out the company's position on open-weights AI models, entering a fast-moving policy debate over whether the United States should restrict how American companies use such systems developed in China. The intervention matters because it comes from one of the leading developers of frontier AI, and because the question sits at the intersection of national security, commercial competition, and the broader direction of AI development.

According to the post, discussion of open-weights models has intensified over the past several days, with particular focus on those originating in China. Amodei notes that reports suggest some US officials are considering banning the use of Chinese open-weights models by American companies. His statement is framed as a response to that emerging debate rather than a reaction to any finalized rule, and it appears to weigh the tension between AI safety, security concerns, and preserving competitiveness.

Open-weights models are AI systems whose trained parameters—the numerical weights that encode what a model has learned—are made publicly available for download. This allows developers to run, fine-tune, and build on the models using their own hardware, without routing requests through the original developer's servers. That is distinct from fully open-source software, which would also include training code and data, and from closed or proprietary models such as Anthropic's own Claude, which are generally accessed through an API. The distinction is central to the policy question, because restricting downloadable weights raises different practical and enforcement issues than restricting a hosted service.

The debate has grown more prominent as Chinese laboratories have released increasingly capable open-weights systems. Models from developers such as DeepSeek and Alibaba's Qwen family have drawn attention for offering strong performance at comparatively low cost, prompting both enthusiasm among developers and concern among policymakers about reliance on foreign-developed AI. Advocates of open weights argue they broaden access, support academic and independent research, and reduce dependence on a small number of vendors. Critics point to questions about security, data handling, potential for misuse, and the strategic implications of widely adopting models built abroad.

Anthropic's involvement is worth noting because the company primarily develops closed models and competes commercially with both open-weights providers and other closed-model labs. That commercial context is relevant when reading any position it takes on the availability of rival systems. At the same time, Amodei has previously written at length about US-China AI competition, including public arguments in favor of maintaining a US and allied lead and support for export controls on advanced semiconductors. Placing this post alongside those earlier essays suggests a consistent focus on the geopolitical dimension of AI, even as the specific recommendations here address open weights rather than hardware.

The wider industry backdrop includes a growing ecosystem of open-weights releases beyond China, such as Meta's Llama models in the United States and Mistral's systems in Europe, which many companies use as a foundation for their own applications. A ban targeting Chinese models specifically would therefore intersect with a market in which open weights of varying origins have become common building blocks. It would also connect to existing US measures aimed at limiting China's access to cutting-edge chips and the tools to make them, extending questions of technology policy from the hardware layer toward the software and model layer.

Practical questions surround any such restriction. Open weights, once released publicly, can be copied and redistributed widely, which complicates enforcement and raises the issue of how a ban on their use would be defined and monitored. Policymakers would likely need to weigh the security rationale against the disruption to developers who already rely on these models, as well as the signal such a move would send about the openness of the broader AI field.

As the source excerpt is truncated, the full detail of Anthropic's specific recommendations is not captured here, and readers should consult the original post for its complete argument. What is clear is that a major AI developer has chosen to weigh in publicly at a moment when the US government is reportedly considering concrete action, adding a prominent voice to a policy question that is likely to shape how open models are treated in the months ahead.