AI’s open model divide

Meta is making another bet that the future of artificial intelligence should not remain entirely behind corporate walls.

The company this week released Muse Glimmer, an open-weight model that developers can download and adapt themselves. The approach differs from the “closed” systems that dominate much of the industry, including the most advanced models from OpenAI and Anthropic, where companies retain control of the underlying technology and users access it through their services. 

The technological divide between the two camps is narrowing. Stanford University's 2026 AI Index found that the leading closed model outperformed the leading open model by 3.3% as of March, with six of the world's 10 highest-ranked models remaining closed.

As open models become more capable, Meta's release adds urgency to a debate over who should control advanced AI, with consequences for competition, security and a growing geopolitical contest over whose technology powers AI systems worldwide.

What opening a model actually means

AI models are not simply open or closed. Developers can release different parts of the systems they build while withholding others.

Model weights are the numerical parameters produced during training that determine how a model processes information and generates responses. Releasing those weights allows another developer to download the finished model, operate it independently and modify it for particular purposes without repeatedly connecting to the original company's servers.

Fully open-source AI can go further by making additional components available, including code and information about how a model was trained. The U.S. National Telecommunications and Information Administration's 2024 Dual-Use Foundation Models with Widely Available Model Weights Report describes openness as a spectrum involving weights, code, training data and technical documentation.

That spectrum now cuts across company lines.

Meta has long championed downloadable models. Nvidia released its Nemotron 3 family in December 2025 alongside training datasets and tools, while OpenAI, despite keeping its most advanced systems proprietary, entered the open-weight market in August 2025 with gpt-oss-120b and gpt-oss-20b. Elon Musk's xAI has similarly released model weights and software openly.

The commercial incentive is substantial. Open weights can reduce dependence on a handful of model providers, allowing companies, universities and governments to customize AI themselves and potentially avoid recurring fees charged by proprietary services.

What happens after release

Greater access comes with a loss of control.

Closed-model providers can change safeguards, monitor use or revoke access, while downloadable weights can circulate independently and be modified to remove safeguards.

The 2026 International AI Safety Report, produced by more than 100 experts nominated by governments and international organizations, found that general-purpose AI is already increasing the speed and scale of some cyberattacks, although it has so far primarily accelerated existing methods rather than created fundamentally new ones. 

Those risks become harder for developers to manage once models circulate beyond their control. Anthropic says open weights can strengthen competition and give users greater control, but argues that sufficiently capable models should undergo rigorous safety testing before release.

The appeal of open models therefore stems partly from the same characteristic that complicates their governance. Developers around the world can use them without obtaining permission from the company that created them, but the original company also loses much of its ability to determine how that technology is subsequently used.

Governments choose how open is too open

Regulators are beginning to draw their own boundaries as open-weight models become more capable.

The European Union has formally incorporated the distinction into implementation of its AI Act. General-purpose models released under qualifying free and open-source licenses can receive exemptions from some transparency obligations when their parameters, including weights, architecture and usage information, are publicly available.

Those exemptions stop when the capabilities become sufficiently significant. Open-source models classified as posing systemic risk remain subject to requirements including model evaluations, adversarial testing, risk mitigation, incident reporting and cybersecurity protections.

The United States has taken a more permissive approach while tying open AI more directly to technological competition. The NTIA recommended in 2024 that the federal government continue monitoring open-weight models rather than immediately restricting their availability, concluding that available evidence did not justify broad limits at the time.

U.S. policy has since moved toward promoting the international reach of American AI. The White House's July 2025 AI Action Plan called for exporting American hardware, models, software and standards, while the Commerce Department's 2026 American AI Exports Program allows open-weight models to participate under certain conditions but excludes those developed by companies based in countries of concern.

The race to become the world's AI

Washington's support for spreading American AI technology is closely tied to China's rapid progress.

The Stanford AI Index found that the performance gap between leading American and Chinese models had fallen to 2.7% by March 2026. Chinese developers have also become major participants in open AI. DeepSeek's releases demonstrated how a model developed outside the dominant U.S. laboratories could spread rapidly among researchers and businesses, while Moonshot AI, a Beijing-based AI startup, has released open-weight versions of its Kimi models as they have grown more competitive with leading U.S.-systems.

Meta argues that restricting American open models risks handing that market to Chinese developers. OpenAI has made a similar geopolitical argument despite maintaining proprietary frontier systems, saying its open-weight models can help countries build on American technology rather than alternatives developed by geopolitical competitors.

For countries without major domestic AI developers, the consequences reach beyond the U.S.-China rivalry. Open-weight models can give businesses, universities and governments from Europe to the Middle East greater ability to operate and customize advanced AI without remaining permanently dependent on a foreign provider.

Muse Glimmer is one model in a much larger contest over how AI will spread. As open systems approach their closed competitors in performance, the models that others are free to build upon could prove as influential as the companies developing the world's most powerful AI.