Meta opens the weights on Muse Glimmer, a 30B model for running agents on consumer hardware

Meta on Monday released Muse Glimmer, a 30-billion-parameter open-weight model built to run AI agents locally on consumer hardware. The weights ship under the permissive Apache 2.0 license, and the model is essentially the open version of Muse Spark, Meta's more powerful closed model, which the company debuted in April.

The pitch is that Glimmer handles multi-step agent work on a single consumer GPU in a Mac or PC: calling tools, writing and debugging code, working with files and screenshots, and carrying a task through an extended workflow. It takes text and images, and Meta says it was trained across more than 100 languages. That local multi-step claim is the one to verify with your own hardware before betting a workflow on it.

A local agent as a privacy play

Meta describes Glimmer's target jobs as managing schedules, drafting messages, and organizing files. Tasks like that live deep in personal data. The privacy angle is doing a lot of work in the release: process everything on the user's device instead of the cloud, and you get an agent that never ships your messages and files to a server.

Meta also wants Glimmer "always-on", operating "anywhere, anytime, with or without an internet connection". No benchmarks accompanied the announcement, so for anyone planning to run an agent on their own box, the missing pieces are the model's multi-step task results and its footprint on a real consumer GPU.

The open part, and the closed part

Glimmer lines up with the direction Mark Zuckerberg has already stated. Last year he argued that advanced AI should empower individuals rather than stay concentrated in a few companies, while warning that Meta would be careful about which of its more powerful models it released openly because of safety concerns.

In a letter on Monday, Zuckerberg repeated the framing: distributing superintelligence widely "has the potential to begin a new era of personal empowerment". He listed what Meta's superintelligence would improve, from a personal agent that "will work 24/7 on your behalf" across relationships, health, career and finances, to tools for starting a business or advancing science. His biggest promise: everyone gets free or affordable access to these tools.

Access is not the same as ownership, though. Meta is increasingly separating the models it releases from the ones it keeps under company control. Muse Spark, the stronger model, stays closed-weight. Glimmer, the smaller one, can be downloaded, fine-tuned, and run on your own hardware.

For a developer deciding where to build, Glimmer draws the first visible line in Meta's plan: the part people can own is an Apache 2.0 model that fits consumer GPUs, and the part people can only access is the stronger Spark that Meta holds back. Which side of that line moves next will define what "personal superintelligence" ends up meaning in practice.