Meta is returning to open source: on 10 August 2026 the company released Muse Glimmer – a multimodal AI model with 30 billion parameters under the permissive Apache 2.0 license. It is designed specifically for local, always-on agent workflows and fits on a single consumer GPU.
Small enough for your own machine
Muse Glimmer is distilled from Meta’s larger Muse model and, according to the maker, runs on a single GPU or a Mac – a 30B model fits on a machine with 24 GB of memory. Meta is clearly targeting local execution: no cloud, no ongoing API costs, and data stays on your own device. The model is tuned for tool use (function calling), long tasks and recovery after failures.
Apache 2.0 and day-0 support
The Apache 2.0 license permits commercial use, modification and redistribution. At launch there is support in widely used tools such as transformers, llama.cpp and vLLM, so the model can be dropped straight into existing setups. Use cases range from local agents and function calling through local coding to evaluating other models (“LLM-as-a-judge”).
In benchmarks, Meta positions Muse Glimmer strongly within its size class – compared, for example, to Gemma4-31B and Qwen3.6-27B. Anyone looking to build local AI agents gets an open, practical foundation to work with.
Sources: VentureBeat, Hugging Face.
