Meta Unveils Muse Glimmer: A 30B Open Agent Model That Runs on a Single Consumer GPU

On August 10, Meta released Muse Glimmer, a roughly 30‑billion‑parameter open‑weight model purpose‑built for running AI agents locally. The model, distilled from Meta’s larger Muse Spark system, is a dense transformer with a separate vision encoder, allowing it to handle both text and images while maintaining long context windows and robust tool‑calling capabilities. According to Meta’s research blog and technical analyses from independent ML outlets, Glimmer has been tuned on agent‑style workflows—things like browsing, using APIs, and orchestrating multi‑step tasks—so it can stay active on a user’s machine rather than living exclusively in the cloud. Crucially, Meta ships the weights under a permissive license and has engineered the model to run in 4‑bit or similar compressed formats, making it feasible to deploy on a single high‑end consumer GPU or recent Macs. CEO Mark Zuckerberg paired the launch with a broader “personal AI agent” vision, arguing that powerful but compact models on user devices will be as important as frontier‑scale systems in data centers. For developers and enterprises, Muse Glimmer sharpens the strategic choice between renting intelligence from proprietary clouds and owning a capable, privacy‑preserving agent stack in‑house.