Meta Platforms has launched a new open-weight artificial intelligence model called **Muse Glimmer**, putting the company back at the center of the debate over whether advanced AI should remain controlled through closed platforms or be made available for developers and organizations to run and customize themselves.
- Muse Glimmer Targets AI Agents Running on Personal Computers
- Zuckerberg Says Larger Open Models Are Coming
- Open-Weight AI Is Different From Closed AI Services
- Chinese AI Developers Are Increasing Pressure on U.S. Companies
- Distillation Becomes Part of the Policy Debate
- Cybersecurity Has Added Another Argument for Open Models
- Meta Plans New Governance for Open Model Releases
- U.S. Government Will Not Put Open Models Through Voluntary Tests
- Meta Is Spending Heavily on AI Infrastructure
- Meta Shares Rose as Investors Assessed the New Strategy
- Muse Glimmer Could Make Local AI Agents More Practical
Muse Glimmer is substantially smaller than the largest frontier models. Meta has positioned it for agentic workloads that can run directly on personal computing hardware rather than requiring a large cloud cluster for every task. Mark Zuckerberg said the model has about **30 billion parameters**, while Reuters reported that it is designed to run on a Mac or PC equipped with a single graphics card.
The release arrived alongside a broader argument from Zuckerberg for reducing barriers facing open-weight AI developers in the United States. Meta also plans to release the weights of **Muse Spark 1.2**, its more capable foundation model, suggesting that Muse Glimmer is the beginning of a renewed open-model strategy rather than a one-off experiment.
Muse Glimmer Targets AI Agents Running on Personal Computers
Muse Glimmer is designed around a different objective from the very largest AI models.
Instead of maximizing model size and requiring expensive infrastructure for every inference request, Meta is targeting smaller agentic tasks that can be performed directly on users’ computers.
Agentic AI systems go beyond answering a single prompt. They can break a goal into multiple steps, use software tools, analyze information, generate content, write or inspect code, and continue performing actions until a task is completed.
Read More
Running that type of AI locally can be particularly useful because an agent may need to make many model calls during a single workflow.
A local model can reduce the need to send every intermediate request to a remote data center, potentially lowering latency and cloud-computing costs while allowing organizations to keep more information within their own systems.
Reuters reported that Muse Glimmer can operate on a Mac or PC using a single graphics card, making the model relevant to developers and organizations interested in bringing AI agents closer to end-user hardware.
Zuckerberg Says Larger Open Models Are Coming
Muse Glimmer will not be Meta’s only new open-weight release.
Zuckerberg said Meta has larger models coming and confirmed that the company intends to release the weights of **Muse Spark 1.2**, its latest foundation model.
Muse Spark 1.2 is also the model behind Meta’s recently introduced **Muse Code** coding system, which is designed to help developers write and debug software, tackle long coding projects, and run multiple sub-agents in parallel.
The decision to open the weights of Muse Spark 1.2 is strategically significant because Meta had previously been one of the most prominent U.S. supporters of openly available AI models.
That position became less certain after the disappointing reception to Llama 4, which contributed to a change in Meta’s AI strategy.
Meta subsequently created a costly superintelligence team as it attempted to regain momentum in the increasingly competitive AI race.
Open-Weight AI Is Different From Closed AI Services
An open-weight model makes the model’s trained weights available for others to download and operate under the terms of its license.
That does not necessarily mean every part of the training process, dataset or development pipeline is completely open, which is why the term “open-weight” is often more precise than describing such systems as fully open source.
For developers and businesses, accessible model weights can provide significant advantages.
Organizations can deploy the model on their own infrastructure, adapt it for specialized tasks, integrate it with internal systems and potentially reduce dependence on a single cloud AI provider.
By comparison, leading models from companies such as OpenAI, Anthropic and Google’s most advanced commercial systems generally operate as closed services whose core weights are not publicly distributed.
Reuters noted that open-weight models are often less expensive to operate than leading frontier systems and can be easier to customize.
Chinese AI Developers Are Increasing Pressure on U.S. Companies
Zuckerberg’s push for a more supportive U.S. policy environment comes as Chinese AI developers have become increasingly competitive in open-weight models.
Reuters highlighted **Moonshot’s Kimi K3, Alibaba’s Qwen3.8-Max and DeepSeek’s V4-Flash** as Chinese models capable of competing with leading U.S. systems.
That competition has changed the strategic importance of open models.
If capable Chinese models are widely downloadable while major U.S. systems remain available primarily through closed platforms, developers around the world can increasingly build products around Chinese model ecosystems.
Zuckerberg argues that U.S. companies should not face unnecessary disadvantages when attempting to develop and distribute open models.
He said American laboratories must comply with additional restrictions involving training data and argued that U.S. policy should reduce that friction if American open-source AI is to remain competitive over time.
Distillation Becomes Part of the Policy Debate
One area highlighted by Zuckerberg is **model distillation**.
Distillation allows developers to use outputs or capabilities from a powerful model to help train a smaller system capable of performing some of the same tasks with significantly less computing power.
The approach is particularly relevant to models such as Muse Glimmer because the AI industry increasingly wants smaller systems that can handle specialized workloads without requiring frontier-scale infrastructure.
Meta wants U.S. policy to provide enough flexibility around training data and distillation for domestic developers to compete with foreign laboratories.
The debate is likely to become increasingly important as AI development splits between extremely large frontier systems and smaller models optimized for devices, enterprises and specialized agents.
Cybersecurity Has Added Another Argument for Open Models
Open-weight AI is also attracting attention because of cybersecurity.
Reuters reported that organizations have become more concerned about recent incidents involving autonomous or highly capable AI systems from several leading laboratories.
One example involved Hugging Face, which said it used a Chinese open-weight model while responding to an attack because some closed models placed restrictions on cybersecurity use.
This illustrates a trade-off in AI safety policies.
Restrictions imposed by closed-model providers can prevent harmful use, but they can also limit legitimate security researchers when they need models for defensive work.
An organization operating its own open-weight model has greater control over how that model is configured and used, although that independence also transfers more responsibility for security and safety to the organization running it.
Meta Plans New Governance for Open Model Releases
Zuckerberg said Meta would establish a governance structure that gives independent directors authority to approve safety criteria for model releases.
The move is intended to address the tension between making powerful AI systems more widely available and managing risks from increasingly capable models.
The policy question has become more complicated as AI models gain abilities related to coding, autonomous tool use and cybersecurity.
Meta’s position is that open availability and safety oversight do not necessarily have to be mutually exclusive.
The company is instead proposing internal governance around release criteria while continuing to make model weights available when those criteria are met.
U.S. Government Will Not Put Open Models Through Voluntary Tests
The policy environment is also changing in Washington.
According to Reuters, the U.S. administration told AI developers earlier in August that open-weight models would not be put through voluntary government safety testing.
The decision matters because a government testing requirement could have added additional time and compliance costs before a company released model weights publicly.
Zuckerberg is simultaneously arguing for fewer restrictions affecting U.S. open-model development, including rules involving training data and infrastructure.
He also rejected the idea that restricting access to foreign open-source models would be an effective response to international competition.
Meta Is Spending Heavily on AI Infrastructure
Meta’s support for AI that can run locally does not mean the company is reducing its investment in gigantic data centers.
Zuckerberg said Meta could spend as much as **$145 billion this year on AI infrastructure**, highlighting the extraordinary capital required to train and operate the most advanced systems.
The company also announced a **$1 billion fund** intended to support communities affected by Meta’s data-center expansion.
Local opposition to data centers has become an increasingly important issue in the United States because large facilities can require substantial amounts of electricity, land, water and new transmission infrastructure.
Zuckerberg argued that the United States is at a disadvantage compared with countries such as China because building infrastructure in the U.S. can be more difficult.
The issue links the open-model debate with a much larger competition over physical AI infrastructure.
Meta Shares Rose as Investors Assessed the New Strategy
Meta shares had fallen roughly **10% for the year** before the announcement.
Reuters initially reported that the stock rose in premarket trading as investors evaluated the new model launch and Meta’s broader AI strategy.
The market reaction comes as investors continue to scrutinize the enormous amounts of money being spent by Meta and other technology companies on AI infrastructure.
For Meta, the challenge is not simply producing technically competitive models.
The company also needs to demonstrate that the billions of dollars being spent on infrastructure, researchers and superintelligence development can create useful products and sustainable economic value.
Muse Glimmer Could Make Local AI Agents More Practical
Muse Glimmer is important not because it is intended to be the largest model available, but because it reflects growing interest in smaller AI systems that can perform real work close to the user.
A 30-billion-parameter model that can operate on suitable personal hardware creates possibilities for coding agents, business automation, research systems, private document processing and other workloads where organizations may prefer not to depend on a remote API for every action.
That does not eliminate cloud AI.
The more likely architecture for many applications is a hybrid model in which smaller local systems handle routine operations while larger frontier models are called when a task requires more advanced reasoning or broader capabilities.
Meta’s decision to release Muse Glimmer and prepare an open-weight release of Muse Spark 1.2 puts the company firmly back into the competition over who will control the software layer underlying that emerging AI-agent ecosystem.
