Muse Glimmer Hardware Requirements: What It Takes to Run a 30B Agent Locally

Seventeen gigabytes. That is what Meta Muse Glimmer weighs once it is quantised to roughly 4 bits, and it is the only number that decides whether the model runs on your machine or does not.

Meta released it on 10 August: 30 billion parameters, Apache 2.0, free to download, and built for local agents rather than chat. It handles text and images, and it is designed to call tools and work through multi-step tasks against your own files, code and screenshots.

Muse Glimmer hardware requirements, in plain numbers

At full precision the model needs over 55GB of memory. No single consumer graphics card comes close, so unless you are running multiple cards or renting a server, full precision is not the version you will use.

Quantised to about 4 bits it compresses to roughly 17GB. That fits on a 24GB card with headroom for a reasonable context window, and sits comfortably on 32GB. Meta shipped quantised builds sized for exactly those two targets, which tells you who they think the audience is.

The general rule holds here: multiply parameters in billions by about 0.6 to get the gigabytes needed at 4 bit. Thirty billion times 0.6 is eighteen, and the real figure lands just under that. Our VRAM tier list works through the same sum for other model sizes.

What to buy if you are starting from nothing

On the PC side, 24GB is the entry point and 32GB is the comfortable one. A card like the ASUS ROG Astral RTX 5090 with 32GB leaves room for a long context window and a second smaller model loaded alongside, which matters more for agent work than it does for chat.

On Apple silicon, unified memory does the same job and the whole pool is available to the model. A Mac mini M4 with 24GB of unified memory clears the 17GB requirement, and our guide to running a local LLM on a Mac covers the setup end to end.

Which runtime to use

The easy path is Ollama, LM Studio or Unsloth. All three will pull a quantised build and give you something usable without much configuration.

For on-device work, llama.cpp and ExecuTorch are the options. If you are serving the model to more than yourself, vLLM and SGLang handle batching and concurrency properly, which the desktop runtimes do not.

Why an agent model is a different purchase to a chat model

A chatbot behind an API sends your prompt somewhere else. An always-on agent that reads your files, watches your screenshots and edits your code does not have to, and that is the actual argument for spending money on memory.

Muse Glimmer is the first open weights release aimed explicitly at that job, at a size that fits a machine a person might own. Whether the tool calling holds up over long multi-step tasks is the open question, and it is worth knowing how model provenance is traced before you build anything important on top of open weights.

If your machine has 24GB, you can run this today. If it has 16GB, you cannot, and no amount of quantisation gets you there without losing something you will notice.

Some links in this article are affiliate links. We may earn a small commission at no extra cost to you.