Meta’s Muse Glimmer puts fully local AI agents on your PC

On August 11, Meta quietly but significantly raised the stakes in the open‑model race with the release of Muse Glimmer, a 30‑billion‑parameter language model whose weights are fully open under the Apache 2.0 license. Unlike many recent “open” models tuned mainly for chat, Glimmer is explicitly optimized for agentic use cases: multi‑step software engineering tasks, complex tool use, and autonomous workflows that can run end‑to‑end on a user’s own hardware.

Meta says the model can run on a single consumer‑grade GPU or modern MacBook, while matching or beating other models of similar size on coding and multi‑step reasoning benchmarks. The company positioned Glimmer as part of a broader push back into open AI, pairing the launch with an essay from CEO Mark Zuckerberg arguing that superintelligent systems should not be locked up inside a few corporations. Meta also hinted at plans to open up a larger successor, Muse Spark 1.2. If Glimmer performs as advertised, it could accelerate a shift toward fully local AI agents—reducing dependence on cloud APIs and making powerful automation accessible even in settings with strict data‑residency or weak connectivity.