Meta on Monday introduced Muse Glimmer, a 30-billion-parameter, open-weight model intended to run multi-step AI agents locally on consumer hardware — a Mac or PC with a single consumer GPU. The company describes Glimmer as essentially an open version of its more powerful closed model, Muse Spark, which Meta debuted in April.
License, inputs and language coverage
Glimmer’s weights are available under the permissive Apache 2.0 license, allowing developers to download and modify them. The model supports both text and images and, according to Meta, was trained across more than 100 languages.
Capabilities: multi-step workflows and on-device data processing
Glimmer is designed to power agents that can perform multi-step tasks such as calling tools, writing and debugging code, working with files and screenshots, and executing extended workflows. By processing information on the user’s device rather than sending it to the cloud, Meta positions Glimmer as a building block for more privacy-sensitive personal agents.
Meta also says Glimmer is designed to be “always-on” and able to function “anywhere, anytime, with or without an internet connection,” enabling continuous local agent services.
Zuckerberg’s vision of personal superintelligence
The release aligns with CEO Mark Zuckerberg’s stated vision of personal superintelligence. Last year he argued that advanced AI should empower individuals rather than remain concentrated in a few companies’ hands, while also warning that Meta must be cautious about which increasingly powerful models it releases openly for safety reasons.
In a new letter published Monday, Zuckerberg reiterated that distributing superintelligence widely “has the potential to begin a new era of personal empowerment,” enabling people to pursue interests, improve their lives, and access tools to start businesses or advance science. He listed areas where such intelligence could assist, from relationships and health to career, finances, home management and hobbies.
Where Meta draws the line between open and closed models
Despite the stated aim of broad access, Meta differentiates between models it will release openly and those it will keep under control. Muse Spark remains closed-weight, while the smaller Glimmer is available for download, fine-tuning, and local deployment on users’ hardware.
Glimmer therefore offers an early indication of how Meta may define the boundary between AI it expects people to own and run themselves and more powerful systems it will retain under company control.
