Muse Glimmer: The Open-Weight Local Agent Meta Just Dropped

Muse Glimmer is a 30-billion-parameter agentic model designed to run locally, with open weights released under the Apache 2.0 license.

Muse Glimmer is a 30-billion-parameter model designed for agents running directly on a Mac or a PC equipped with a capable GPU. Its open weights are released under the Apache 2.0 license, with a focus on persistent personal assistants, local coding, tool calling, and autonomous workflows.

The model accepts both text and images through a dedicated perception encoder. It can chain tool calls, maintain a plan across multiple steps, recover after a failed action, and work with different orchestration environments. Meta also says the model was trained on data spanning more than 100 languages and supports multiple reasoning effort levels.

To make local inference more accessible, the weights have been quantized to approximately 4-bit precision, bringing the language model portion below 20 GB. A K-Quant version takes up around 17 GB, leaving memory available for context, image understanding, and the auxiliary model used to accelerate generation. Meta says it tested the full setup on systems with 24 GB or 32 GB of memory.

Part of that acceleration comes from a lightweight DFlash model that proposes multiple tokens ahead of time, which Muse Glimmer then verifies in parallel. The company reports speeds suitable for real-time agent interactions on MacBook M4 Max and M5 Max systems, as well as an RTX 5090.

Training combines distillation from Muse Spark, long-context data, supervised fine-tuning, on-policy distillation, and reinforcement learning. In one demonstration, Muse Glimmer discovers a Home Assistant instance on a local network, queries its devices, builds a web interface, and then starts a local server to verify it.

The open weights are available on Hugging Face. Optimized integrations for llama.cpp, MLX, and ExecuTorch are expected in the coming days, along with support through Ollama, LM Studio, Unsloth, vLLM, SGLang, and several inference providers.