Skip to content

Meta releases Muse Glimmer, 30B open-weight model for local agents

devAug 10, 202690192

Meta released Muse Glimmer, a 30-billion-parameter open-weights model engineered to run locally on a single laptop or GPU and downloadable for offline use. Meta published the weights under an open license and designed Muse Glimmer for agentic workflows with built-in tool use, multimodal inputs including vision, and long context support. Independent reporting and Meta posts say the model’s quantized weights can fit under about 20 gigabytes of VRAM, enabling local inference on mainstream consumer hardware. Meta said it will follow with an open-weights release of its more advanced Muse Spark 1.2 in the coming weeks, and CEO Mark Zuckerberg urged lower U.S. barriers for open-source models in the same announcement. Meta positioned the release as part of a push to let developers run, adapt, and test models offline for privacy, customization, and edge deployments. Early community benchmarks place Muse Glimmer broadly comparable to other dense 27, 31B-class models on agentic tasks while independent testers continue to evaluate long-run agent stability and prompt-injection resistance.

3 sources