Meta Muse Glimmer runs a 30B local agent on one high-memory GPU

Meta's Muse Glimmer is a 30-billion-parameter, Apache 2.0 model built to run agent workflows on local hardware. Its official model page and Hugging Face model card say the 4-bit language model is under 20 GB, while the complete setup targets machines with 24 or 32 GB of memory.

That hardware requirement is the catch. NVIDIA demonstrates the model on an RTX 5090, while AMD's test setup assigns 64 GB of variable graphics memory on a 128 GB Ryzen AI Max system. This is practical local AI for a high-end workstation, not a typical laptop upgrade.

Muse Glimmer can take text and images, call tools, recover from failed steps, and work across a long context window. Meta reports scores of 75.5 on MCP Atlas and 51.2 on SWE-Bench Pro, but those remain vendor evaluations rather than independent proof. The model card also recommends guardrails and human confirmation before irreversible actions. The launch was reported by Reuters on August 10.