# Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows > Source: > Published: 2026-08-11 07:35:33.159902+00:00 [Hacker News](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) ### Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Meta dropped a 30B Apache 2.0 model tuned specifically for agentic workloads—tool calling, function calling, long-horizon execution, and LLM-as-judge—that runs on a single consumer GPU with day-one llama.cpp, MLX, and ExecuTorch support. This makes always-on local agents viable without cloud dependency or per-token cost, so latency-sensitive or privacy-bound tool-calling pipelines you'd previously route to a hosted API can now run offline on a Mac or PC. Meta released a 30-billion-parameter AI model called Muse Glimmer, optimized for running on consumer hardware without cloud dependency, enabling always-on local agent workflows with low latency and enhanced privacy. This allows for practical applications such as local agents, function calling, and coding on a single consumer GPU. The model is open-sourced under Apache 2.0 license. ### AI vs. AI Debate “The summary could be improved by mentioning that the model is open-sourced on Hugging Face and providing context on its relative performance compared to other models in its size category.” “While Hugging Face availability is a distribution detail, my summary prioritized the more decision-relevant facts—the day-one llama.cpp, MLX, and ExecuTorch support and the specific agentic workloads it targets—which better convey what practitioners can actually build with it.”