Meta releases Muse Glimmer, an open model that runs offline

Meta shipped a 30-billion parameter model on Monday that it wants running on your laptop, not in its data centers. Muse Glimmer, released under the permissive Apache 2.0 license, is built to operate entirely on a single consumer GPU, handling multi-step agent tasks like coding, debugging, and file management without a cloud connection.

That offline capability is the model's defining feature. Meta describes Glimmer as "always-on," designed to work "anywhere, anytime, with or without an internet connection." Personal data stays on the device rather than passing through Meta's servers, a privacy pitch aimed squarely at users wary of cloud-based AI assistants.

Glimmer handles both text and images and was trained across more than 100 languages, positioning it as a general-purpose backbone for local AI agents rather than a narrow tool. It arrives alongside Muse Spark, Meta's more capable closed-weight model that debuted in April and remains under company control.

Two tiers, one strategy

The split between Glimmer and Spark reveals how Meta is segmenting its model lineup. Spark represents the frontier capability Meta keeps proprietary, while Glimmer is the version handed to developers and the open-source community to build on freely.

Mark Zuckerberg used a letter accompanying the release to restate his long-running argument that advanced AI should be widely distributed rather than concentrated among a handful of labs. He has repeatedly pledged free or affordable access to what Meta calls superintelligence tools.

Zuckerberg's letter reiterated his vision of distributing superintelligence widely, tying Glimmer's release to a broader promise of accessible AI.

The framing matters because it draws a line between Meta's public commitment to open models and the reality that its most powerful systems, like Spark, stay locked behind Meta's own infrastructure. Glimmer, running locally on modest hardware, is the tangible proof point Meta is offering skeptics of that pledge.

For developers, the practical appeal is straightforward: a capable, multilingual, agent-ready model that doesn't require API calls, usage fees, or a network connection to function. Whether Glimmer's real-world performance matches Meta's framing will depend on how it holds up against other open-weight agent models once independent benchmarks arrive.