Meta open-sourced Muse Glimmer 30B, aimed at local agents. 4-bit, it fits under 20GB, small enough to run on a single consumer GPU. Tool calling, multi-step planning, 137 languages: it beats Gemma4-31B and Qwen3.6-27B on each. Apache 2.0. The model and the agent both live on your own box now, and for once that's enough to do real work.
Meta open-sourced Muse Glimmer 30B, aimed at local agents.