Why open-weight models are crucial for American AI leadership

The greatest single potential threat to sustained American artificial intelligence (AI) leadership would be new policies that impede the domestic development of a thriving ecosystem of open-weight AI providers and solutions. To see why, it is helpful to explain what open-weight models are, how they fit within the broader context of AI, and why they are so important for the digital economy.

A spectrum of AI systems

AI systems exist along a spectrum. On one end are closed models, such as the various tiers of Anthropic’s Claude and OpenAI’s ChatGPT. Users can submit queries to these models and get responses, but the detailed inner workings of the models and how they were trained are proprietary.

At the other end of the spectrum are open-source models, in which information about the training data, the code used to generate weights from the training data, and the resulting weights are publicly disclosed. In between these two endpoints are “open-weight” models in which the weights that “determine how the model interprets input data and generates outputs” are public, but the methods that were used to generate those weights are not.

The importance of a diversity of model types

An AI ecosystem that lacks robustness in any of these AI categories is deficient, as each type of AI plays a distinct and complementary role. Providers of closed models retain full control over how their models are deployed and used, giving them a strong incentive to invest heavily in providing a continuously updated, highly curated set of products that operate within carefully constructed guardrails.

Open-source models are invaluable for researchers and companies seeking a foundation that provides the maximum degree of latitude in developing enhanced, customized in-house solutions. Open-weight models are vital for users who need the flexibility to use or adapt a set of pre-computed model weights but would gain little additional benefit from access to the full set of information about how those weights were derived. Both open-source and open-weight models are important tools to help the next generation of AI researchers and entrepreneurs learn about AI. Solutions built on open models can also be much more cost-effective over the longer term than solutions that rely on continual payments to providers of closed models.

Open-weight models are not inherently less secure

Discussions about open-weight AI models often raise issues of risks including security. But open systems can actually be the foundation for building more secure systems, as they are open to scrutiny by anyone and easier to evaluate. This contrasts with closed models, which have inner workings examined only by the people within the companies that developed them.

Open-weight models are also criticized because they lack the guardrails that are found in closed models. This is a fair critique, but there is a mirroring critique that cuts the other way: The guardrails built into closed models can create their own cybersecurity concerns, because they can make it impossible to use the model to identify and patch security holes in third-party systems.

Consider the inadvertent July 2026 intrusion into Hugging Face’s infrastructure by a pair of OpenAI models that escaped their sandbox. In a post describing how it analyzed the intrusion, Hugging Face explained that the “models we reached for first [to help investigate], Claude Opus and Fable, refused a large part of that work: their safety guardrails treated reverse-engineering an exploit the same as launching one.” Having been unable to make sufficient progress using closed models, Hugging Face then turned to Nvidia’s open weight GLM-5.2 NVFP4 model, which is in turn a “quantized” (i.e., streamlined so that runs faster but at potentially lower precision) version of ZAI GLM-5.2, an open-weight model from Chinese company Z.ai. The open-weight model provided valuable information in enabling Hugging Face to unravel how the intrusion occurred.

Open-weight models can run locally, with less delay, and more reliability

Closed AI systems are run on the cloud. This involves a delay as information is sent from an “edge” device to the cloud and back again. For some applications this delay—for example, a few tens of milliseconds—is inconsequential. But for many applications, such as navigating an electric vehicle (EV) or a search and rescue drone, that delay is unacceptable.

By contrast, open-weight models (and more specialized systems derived from them) can be run locally on “edge” devices, thereby eliminating the delays inherent in relying on remote computing infrastructure. Running models locally also removes the potential operational impediments posed by outages in the cloud or in the communication links between the cloud and edge device.

Open-weight models promote a robust ‘harness’ ecosystem

While much of the discussion about AI focuses on models, that isn’t the full story. How an AI model operates in practice is also determined by the particular scaffolding and harness that surrounds it. An example is Claude Code. Claude is an underlying family of models that are useful for an enormous spectrum of tasks. Claude Code adds a harness to this framework to create a product that is optimized for producing computer code.

While Claude Code is extremely good, it is a harness built on a closed system. Providers of closed systems will be limited in how many harnesses they will build for their models. By contrast, open-weight systems allow anyone to develop (and potentially monetize) a harness to optimize the model for use in a specific task. More harnesses mean that AI can be useful to more people and for a greater variety of applications.

Banning Chinese open-weight models is the wrong solution

The best American response to robust overseas competition is to innovate faster. Bans lower the incentives to innovate and invite retaliation from other nations. The market for EVs, where punitive U.S. tariffs on Chinese EVs amount to a de facto ban, furnishes a useful analogy. Protectionist U.S. EV policies have resulted in higher prices and less choice for American EV buyers, less EV innovation in the U.S., and the rapid acceleration of market dominance by Chinese EVs in countries outside the U.S.

While calls in U.S. policy circles to place restrictions on Chinese open-weight AI models typically cite (with little to no supporting evidence) alleged threats to national security, that sidesteps the fact that Chinese open-weight models are extraordinarily good. Unless artificial trade barriers intervene, people around the world will seek out and adopt the best models, regardless of their country of origin. If U.S. companies can build open-weight models that are on par with or better than the best Chinese models, the U.S. models will experience wide global adoption.

Developing the business case for open-weight models

Startups in the open-weight space include Poolside, Reflection, and Hugging Face in the U.S.; DeepSeek, Moonshot AI, Z.ai (also called Zhipu AI) in China; and Mistral in France. These and other companies in the field are using a variety of approaches to generate revenue, with offerings including customer-specific model adaptation, hosting services, and agents that “plan and execute multi-step tasks” within an enterprise.

To date, venture capital investments and valuations in the open-weight space are far smaller than with the leading closed model providers, OpenAI and Anthropic. But AI changes fast, and while the analogy is imperfect, Linux and Android provide important precedents showing that open systems can be the foundation for highly profitable products and services.

Open-weight models are foundational to democratizing computation and AI

If American AI is dominated by a small handful of closed model providers, the result will be an overconcentration of power. That raises many concerns, including the risk of regulatory capture and a set of resulting policies that impede new market entrants. A far better approach is to foster an environment in which closed model providers and open model providers can both thrive, offering complementary products and services.

Computing in the age of AI is creating extraordinary opportunities for innovation across fields including health care delivery, education, pharmaceuticals, manufacturing, and many more. The best way to allow 21st century American ingenuity, amplified by AI, to reach its full potential is to promote growth of a diversified ecosystem with global leadership across all categories of AI.