Meta Releases Muse Glimmer — an Open Agentic Model with 30 Billion Parameters

Meta's Superintelligence Labs division has announced the release of a new large language model, Muse Glimmer. It has 30 billion parameters and is intended for always-on local agentic scenarios. The model's weights are published under the Apache 2.0 license, meaning they are distributed as open source.

By the developers' design, Muse Glimmer is compact enough to run on a single consumer graphics accelerator. That makes it suitable for local agents that help write code, for scoring responses under the “LLM-as-a-judge” scheme, and for other tasks. The model was trained and tested on end-to-end execution of agentic tasks and multi-step reasoning, and it was also optimized for local deployment.

Details on Muse Glimmer are available at research.meta.ai, and the model itself can be downloaded from Hugging Face. A number of open projects have already added support for the newcomer — in particular, the tools updated that same day included ollama version 0.32.7.

The bet on local execution and an open license continues an emerging trend: instead of gigantic cloud models available only through an API, vendors are increasingly offering compact variants that users can keep “at hand” on their own hardware and run without a constant connection to external services.

Source: Phoronix