Muse Glimmer

Meta returned to open weights in August 2026 with Muse Glimmer, a roughly 30B parameter dense transformer distilled from its closed Muse Spark system and licensed Apache 2.0. The 52-layer network mixes sliding-window and global attention in a repeating three-local-one-global pattern with 2,048 token windows, carries a 1.8B parameter vision encoder for image input, covers more than 100 languages, and reads 131K tokens of context. Training centered on end-to-end agent behavior: reliable function calling, multi-step tool workflows, error diagnosis, and recovery from failed actions. What makes it notable is deployment reach rather than raw scale; official K-Quant variants compress the language model below 20GB at 4-bit precision so it runs on a single 24GB consumer card or an Apple Silicon MacBook, and the bundled DFlash drafter adds block-level speculative decoding measured at 3.1x on an RTX 5090 and 1.5x on an Apple M4 Max. That combination puts a full agent loop, including vision, on hardware hobbyists already own, with no usage restrictions attached.