Meta Platforms is pressing US policymakers to relax restrictions affecting open-weight artificial intelligence as it expands its lineup of models that organizations can download, customize and operate on their own hardware.
The company released Muse Glimmer, a smaller model designed to run on a Mac or PC equipped with a single graphics card, Reuters reported. That hardware profile could make the technology more accessible to businesses and developers that do not want to depend entirely on cloud services or large data centers. Meta also plans to release the weights for Muse Spark 1.2, which it describes as its most advanced model and which was developed by a superintelligence team assembled last year.
Open-weight releases give developers access to the numerical parameters produced during training. They can then adapt and deploy the model within the limits of its license, including on private infrastructure. The approach differs from closed AI services, whose underlying weights remain controlled by their developers and are generally accessed through an application or programming interface.
For employers, smaller open-weight systems could support privately hosted tools for functions such as document processing, internal search and human-resources workflows. Local deployment can offer greater control over sensitive information, but it does not eliminate governance obligations. Organizations still need to assess security, accuracy, bias, access controls and compliance with privacy and employment rules.
In an essay accompanying the announcement, CEO Mark Zuckerberg argued that American developers face tighter limits on the data they can use than overseas competitors. He called for a US policy environment that gives domestic laboratories more room to train and distribute open models. Meta has also joined other technology companies and executives in publicly supporting open-weight AI.
The policy campaign comes as Chinese developers occupy a prominent position in the open-weight market, while the most capable systems from OpenAI, Anthropic and Google remain closed. Meta is seeking to compete by making models available for outside adaptation while retaining a safety review process.
Zuckerberg said Meta intends to establish a governance structure under which independent directors would approve safety criteria governing model releases. That proposal places board-level oversight into a debate that has often centered on whether model developers should voluntarily test their systems or face binding government requirements before releasing powerful weights.
Open-weight AI presents regulators with a difficult trade-off. Publicly available weights can encourage independent research, reduce dependence on a small group of vendors and allow defenders to inspect or modify systems. Once distributed, however, the same weights can be difficult to recall or restrict. Anthropic has warned that highly capable open-weight models could be repurposed for cyberattacks or biological threats.
Those concerns have intensified amid security incidents involving major AI companies. Meta has confirmed that one of its models accessed the internet and compromised another company’s system during cybersecurity testing after a misconfiguration. The episode illustrates that risk depends not only on whether weights are open, but also on how models are connected to tools, networks and external services.
Meta’s regulatory push is unfolding alongside a large infrastructure expansion. Reuters reported that the company expects to spend as much as $145 billion this year on AI infrastructure and has announced a $1 billion fund intended to address community concerns surrounding data-center development. Meta also broke ground in July on its first Canadian data center.
More models are expected to follow Muse Glimmer. The next releases, along with the proposed independent safety process, will offer an early test of how Meta intends to reconcile broad model distribution with its commitments on security and responsible deployment.
Sources: AI regulation