Meta's Muse Glimmer does something no frontier-adjacent model has done this cleanly: the 4-bit build is an 18GB download that runs on one consumer GPU or a MacBook, with no account, no API key and no metered tokens. If your machine has 24GB of VRAM or a recent Apple Silicon chip, you can be talking to it in under ten minutes. This guide covers the three sane ways to install it, from easiest to most control.
What you need first
- 24GB VRAM GPU (RTX 3090/4090/5090 class) for the standard 4-bit quant, which needs about 17GB. With 32GB you can run the higher-quality K-Quant-Dynamic build, which Meta says degrades output by only 0.2% versus full precision.
- Or a Mac with Apple Silicon — 32GB unified memory is comfortable, 24GB works with the smaller quant. Meta's own numbers show an M4 Max doing 37.8 tokens/sec with speculative decoding, and an M5 Max hitting 50.2.
- Or 64GB VRAM if you want full-precision BF16 — that's workstation territory, and for local agent work the 4-bit build is the one to run anyway.
- Around 20-25GB of free disk per variant.
The model handles text and images, supports a 131,072-token context window, and ships under Apache 2.0 — commercial use included, no strings.
Option 1: Ollama — the ten-minute path
Ollama already has the model. If Ollama is installed, this is the whole tutorial:
ollama run muse-glimmer
That pulls the 18GB 4-bit build with the 128K context window. On a Mac, use the MLX-optimized variant instead — it runs on Apple's MLX engine and is noticeably faster on M-series chips:
ollama run muse-glimmer:30b-mlx
First run downloads the weights; after that it starts in seconds. Ollama also exposes an OpenAI-compatible API on localhost:11434, which means any agent framework that can point at a base URL can now use Glimmer as its brain.
Option 2: LM Studio — if you want a GUI
LM Studio lists Muse Glimmer in its model browser (the Hugging Face page counts 62 community quantizations, and LM Studio surfaces the good ones). Search for "Muse Glimmer", pick a quant that fits your RAM — the ~17GB 4-bit for 24GB machines, the K-Quant-Dynamic for 32GB — and click download. You get a chat window, a local server toggle, and none of the terminal. This is the right path if you mostly want to try the model before wiring it into anything.
Option 3: vLLM — if you're serving agents
For a box that will serve the model to multiple agents or users, vLLM is the production path:
pip install vllm
vllm serve "meta-models/Muse-Glimmer-30B"
Python users who want the raw model in code can use transformers directly:
from transformers import AutoProcessor, AutoModelForMultimodalLM
processor = AutoProcessor.from_pretrained("meta-models/Muse-Glimmer-30B")
model = AutoModelForMultimodalLM.from_pretrained(
"meta-models/Muse-Glimmer-30B", device_map="auto")
Note the class name: Glimmer is multimodal, so it loads through AutoModelForMultimodalLM, not the plain causal-LM class you may have in old scripts.
Settings that actually matter
Meta's model card recommends temperature 1.0, top_p 0.95 and top_k 64 — unusual, since most models want lower temperatures for agent work, so don't carry over your old settings by habit. For hard tasks the card suggests setting reasoning strength to "high" or "xhigh".
The other headline feature is the DFlash speculative-decoding drafter that ships alongside the model. With it enabled, Meta measured a 3.1x speedup on an RTX 5090 — 233 tokens/sec against a 75 tok/s baseline. Ollama and LM Studio are expected to wire this in over time; vLLM users can enable it today by following the model card.
Is it worth installing?
If you run coding or automation agents locally, yes — this is the first Apache 2.0 model tuned specifically for agent work (scheduling, file management, tool use) that fits on hardware people actually own. The cloud-model gap has not closed, but for an always-on local agent that costs nothing per token, Glimmer moves the floor up substantially. And since the weights are on your disk, no pricing change, quota cut or deprecation notice can take it away.
The quickest honest test: ollama run muse-glimmer, paste in a real task from your week, and see whether the answer clears your bar. Ten minutes, zero dollars.