Muse Glimmer is a 30-billion-parameter model licensed under Apache 2.0. The model is designed for local agent workflows and runs on a PC or Mac with a single consumer-grade GPU.
The model focuses on what Meta calls “always-on” local agents. Examples include function calling, local programming, and LLM-as-a-judge evaluations. Everything runs on the device itself, without an internet connection.
Muse Glimmer is distilled from Muse Spark, the model that Meta Superintelligence Labs presented in April of this year. Training took place in three phases: logit distillation on the output of the larger model, mid-training with longer context and agentic data, and finally, supervised fine-tuning combined with reinforcement learning.
From 55 GB to under 20 GB
At full precision, a model with 30 billion parameters would require over 55 GB of memory. No consumer GPU offers that much. Meta therefore compresses the weights to approximately 4-bit precision, keeping the language model under 20 GB. This leaves room for the KV cache, the perception encoder for image understanding, and the drafter for speculative decoding, all within 24 or 32 GB.
That drafter is based on DFlash and proposes entire blocks of tokens at once, which the main model verifies in parallel. According to Meta, this results in 3.1 times faster decoding on an RTX 5090. On a MacBook with M5 Max, it’s 1.8 times faster, and on an M4 Max, 1.5 times faster.
The model alternates between processing text and images via a separate perception encoder, enabling agents to interpret screenshots, graphs, and documents. Additionally, it is trained on data from over a hundred languages, and Muse Glimmer supports adjustable reasoning strengths. If a tool call fails, the model is trained to diagnose the error and retry, rather than stopping.
Availability
The model is now available on Hugging Face. Optimized integrations for llama.cpp, MLX, and ExecuTorch will follow in the coming days. Partners such as Ollama, LM Studio, Unsloth, Together AI, Fireworks AI, and OpenRouter are also supported. For scalable inference, Meta points to vLLM and SGLang.
In addition, Meta is collaborating with AMD, Arm, Dell, Intel, and Nvidia on device-specific optimizations. The model was evaluated under Meta’s Advanced AI Scaling Framework before being released as an open-weight model.
Tip: Meta is launching Muse Code: an AI agent for large codebases