Meta Releases Local Muse Glimmer with Clean License for Agent Work

Meta has released Muse Glimmer, a 30B vision model.

Moving Away from Restrictive Llama Licenses

It ships with a clean Apache 2.0 license. The model fits comfortably on machines with 32 GB RAM.

Tuned for Tool Use First

Glimmer isn't just another conversational chatbot. Meta optimized it against benchmarks like DeepSearch QA, SWE-Bench, and MCP-Atlas. It can sustain coherent planning over longer horizons and adheres to strict schemas when invoking local functions without hallucinating syntax.

Local Orchestration Without Data Leaks

Reliable tool use has traditionally required calling large cloud APIs. Packing this capability into a 30B open-weights model under an Apache license lets developers build agents that interact with sensitive local code without sending prompts to third parties.

Long Context as the Main Quality Test

The real test is how often the model loses context after twenty tool-calling steps. Benchmarks look promising, but production deployment inside high-context dev tools will prove if Glimmer can maintain focus without massive cloud scale.

Lilith's verdict

Meta just released a fully functional clerk that will scour your local network without phoning home.