The Facebook parent company on Monday, August 10, released Muse Glimmer, a new artificial intelligence model designed to run AI agents directly on personal computers, while CEO Mark Zuckerberg called on Washington to remove barriers that he says are making it harder for American developers to compete with rapidly advancing Chinese AI companies.

The launch marks a significant change in Meta's AI strategy after the company pulled back from releasing some of its most capable models openly. Meta is now promising not only Glimmer but an open-weight version of the much more powerful Muse Spark 1.2 in the coming weeks.

What Is Muse Glimmer AI?

Muse Glimmer is a 30-billion-parameter open-weight AI model developed by Meta Superintelligence Labs.

Unlike massive frontier models that generally depend on large cloud data centers, Glimmer was designed for local agentic AI workloads.

That means developers can use the model to build AI systems capable of performing multi-step tasks, calling tools, maintaining context and interacting with local applications without having to send every request to a remote AI server.

Reuters reported that Muse Glimmer can run on a Mac or PC using a single graphics card, putting considerably more capable AI within reach of developers who want to operate models on their own hardware.

AMD separately confirmed that Muse Glimmer is a 30B dense model released under the Apache 2.0 license and demonstrated it running on Ryzen AI Max+ systems and a workstation equipped with a single Radeon AI PRO R9700 GPU.

Why Muse Glimmer Running Locally Matters

Most popular generative AI services operate primarily in the cloud.

A user's prompt is transmitted to remote servers, processed by the model and then returned to the device.

Muse Glimmer targets a different model: AI running locally on the user's machine.

This approach could be particularly attractive for developers building coding assistants, research tools, automation software and private enterprise applications.

Local AI can potentially reduce recurring API expenses while giving developers greater control over where information is processed. AMD says local deployment also allows developers to keep model workloads and associated data on the user's machine.

The model is particularly focused on agents rather than simple question-and-answer chatbots.

AMD's technical description says Glimmer is designed to keep context, make sequences of decisions, call tools, examine results, adapt when something goes wrong and continue working across longer workflows.

That places Muse Glimmer squarely in one of the biggest trends in artificial intelligence: agentic AI.

Muse Spark 1.2 Is Coming Next

Glimmer may not be Meta's biggest AI announcement for long.

Meta says it intends to release the weights for a version of Muse Spark 1.2 in the coming weeks. Axios describes Spark as Meta's foundation model, while Reuters reports that Spark 1.2 is the company's most advanced model.

Zuckerberg teased the upcoming models while unveiling his broader vision for AI.

“We've got even bigger models that are coming soon,” Zuckerberg said.

The decision could put Meta back near the center of the open-model ecosystem.

Meta was previously one of the most prominent American supporters of downloadable AI through its Llama models. However, the poor reception of Llama 4 contributed to a strategic shift before the company began rebuilding its AI organization around Meta Superintelligence Labs.

Muse Glimmer now signals that open releases are once again a major part of that strategy.

Zuckerberg Says the US Is Falling Behind China in Open AI

Muse Glimmer's launch also comes with a political message.

Zuckerberg argues that American AI developers face restrictions that could eventually allow foreign competitors — particularly Chinese AI laboratories — to dominate open-weight artificial intelligence.

In his new essay, “The Future is for Everyone,” Zuckerberg called for AI technology to be distributed broadly rather than controlled by a handful of corporations or governments.

He specifically argued that American developers face greater restrictions around areas such as training data and model distillation.

“U.S. policy must reduce this additional friction if we want American open source models to lead over time,” Zuckerberg said.

Model distillation allows developers to use the capabilities or outputs of a larger AI system to train a smaller model that can perform similar tasks with significantly lower computing requirements.

Chinese AI Models Are Raising the Pressure

Meta's renewed open-weight push comes as Chinese AI companies become increasingly influential among developers.

Reuters identified models from DeepSeek, Alibaba and Moonshot AI among Chinese systems competing with leading American AI technology. At the same time, many leading models from major US developers remain closed-weight.

That creates a strategic problem for US technology companies.

An open-weight model can be downloaded, modified and deployed by independent developers. This can encourage a much larger ecosystem around a model than would be possible with a system available only through a company's proprietary cloud service.

Muse Glimmer appears designed to help Meta reclaim that developer audience.

Is Muse Glimmer Really Open Source?

The most technically accurate description is open-weight AI, although Zuckerberg and some reports also use the broader term “open source.”

Muse Glimmer's trained model weights are available to developers, and AMD says the model is licensed under Apache 2.0, allowing developers to use, modify and redistribute it, including for commercial purposes.

That makes Glimmer considerably more permissive than many AI models that can only be accessed through commercial APIs.

The distinction between “open-weight” and “open-source,” however, remains important because an open-weight release does not automatically mean every element involved in creating a model — such as the complete training dataset and training pipeline — is publicly available.

What Hardware Does Muse Glimmer Need?

Muse Glimmer is significantly smaller than the largest frontier AI models, but running a 30-billion-parameter model locally still requires capable hardware.

AMD says supported systems with more than 32GB of VRAM or Variable Graphics Memory can run Muse Glimmer using tools such as LM Studio.

The company recommends systems based on Ryzen AI Max+ processors or its Radeon AI PRO R9700 32GB graphics card for out-of-the-box local deployment.

AMD's preliminary tests reported generation speeds of up to 24 tokens per second on a Ryzen AI Max+ 395 and up to 53 tokens per second on a Radeon AI PRO R9700 under its tested configuration, although actual performance will depend on hardware and software settings.

Meta's Bigger Bet: AI for Everyone

Zuckerberg's argument goes beyond Muse Glimmer.

In a roughly 6,500-word AI manifesto published Monday, he argued that concentrating the most powerful artificial intelligence systems within a small number of governments or corporations could itself create substantial risks.

His preferred alternative is widespread access.

Meta plans to resume releasing selected open AI models while creating governance procedures under which independent directors can approve safety criteria for model releases.

Zuckerberg also announced a $1 billion “Future is for Everyone Fund” aimed at supporting US communities where Meta owns and operates data centers, including spending related to teachers, first responders and local energy and water infrastructure.

Why Muse Glimmer Could Be Important

Muse Glimmer is unlikely to compete directly with the world's largest frontier models purely on size.

That may not be the point.

Meta is betting that developers increasingly want models that are capable enough to perform real work but small enough to operate privately and economically on local hardware.

If that trend accelerates, the competition in AI could move beyond simply asking which company has the smartest chatbot.

The next battle may be over which AI model developers actually choose to download, customize and put inside their own applications.

Muse Glimmer gives Meta a new contender in that race.

And the planned open-weight release of Muse Spark 1.2 could make the competition considerably more intense.

Muse Glimmer AI FAQ

What is Muse Glimmer AI?

Muse Glimmer is a 30-billion-parameter open-weight artificial intelligence model released by Meta Superintelligence Labs and optimized for local agentic AI workloads.

Can Muse Glimmer run on a PC?

Yes. Meta designed Glimmer to run locally on Macs and PCs with sufficiently capable hardware, including systems using a single graphics card.

Is Muse Glimmer free?

The model is released under the Apache 2.0 license, which permits commercial use, modification and redistribution subject to the license's terms.

How large is Muse Glimmer?

Muse Glimmer contains approximately 30 billion parameters.

What is Muse Spark 1.2?

Muse Spark 1.2 is Meta's more advanced foundation AI model. Meta says it plans to open the weights for a version of Spark 1.2 in the coming weeks.

Why is Meta releasing open-weight AI models again?

Meta says wider access to AI can encourage innovation and prevent advanced AI from becoming concentrated among a small number of institutions. Zuckerberg also argues that stronger American open-weight models are necessary for the US to remain competitive with Chinese developers.

What is the difference between Muse Glimmer and ChatGPT?

Muse Glimmer is primarily an open-weight model that developers can download and run on their own compatible hardware. ChatGPT is a hosted AI product. The products therefore target different deployment models even though both involve generative AI.