ONLINEAGENT_OPS 2026.Q3 HOME ARTICLES CRAFT RECORD BLOG MAP HUBS FAQ SEARCH
HOMEARTICLESHOW TO RUN MUSE GLIMMER LOCALLY
ARTICLES · GUIDE

How to Run Muse Glimmer Locally

Meta’s 30B open agent model needs about 24 GB of GPU memory in its quantised form. The published sizes, the routes Meta names, and what is not confirmed yet.

READ4 min
WORDS724
SECTIONS5
TYPEGUIDE
CHECKED16 SEP 26

You need about 24 GB of GPU memory and the quantised weights. Muse Glimmer is a 30-billion-parameter agent model Meta released under Apache 2.0 on 10 August 2026, and it was built to run on one consumer card rather than in a data centre. How it compares with Google’s open family is on Gemma 4 vs Muse Glimmer.Meta AI Research, Introducing Muse Glimmer, 10 Aug 2026, read at source 16 Sep 2026: “Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows.”

TL;DR — THE SHORT VERSION

The size question has a published answer, and it is not the raw parameter count.

  • Full precision needs over 55 GB. Meta’s own words, and no consumer GPU has it.
  • The 4-bit version is under 20 GB, which is what makes a single card possible.
  • Meta targets a 24 GB or 32 GB envelope — weights, context, image encoder and the speed-up drafter together.
  • Apache 2.0. The weights are on Hugging Face, not behind a bespoke licence — a change of position covered on is Llama dead.
  • It is an agent model, not a chatbot. Tool calling, long tasks, and recovering from failed calls are what it was trained for.

What hardware do you need?

Meta publishes the arithmetic rather than a recommendation. “At full precision, a 30-billion parameter model would require over 55 GB of memory — far more than any consumer GPU offers.” The fix is compression: “We use quantization techniques to compress the model’s weights to approximately 4-bit precision, shrinking the language model to under 20 GB.”Meta AI Research, Introducing Muse Glimmer, 10 Aug 2026, read at source 16 Sep 2026.

Under 20 GB is not the whole budget. Meta says the compression leaves room for three more things beside the weights: “the perception encoder for image understanding, and the speculative decoding drafter to run simultaneously within a 24 GB or 32 GB envelope”, alongside the working memory a long context needs.

The machines Meta measured on are named: “MacBook M4-Max, M5-Max and on a RTX-5090.”

MUSE GLIMMER · MEMORY NEEDED
Meta’s own figures from its 10 Aug 2026 announcement, not measured here. Full precision and the 4-bit weights are drawn at their stated bounds.
Full precisionover 55 GB
Envelope Meta targets · larger32 GB
Envelope Meta targets · one 24 GB card24 GB
4-bit weights aloneunder 20 GB
Meta AI Research, Introducing Muse Glimmer, 10 Aug 2026, read at source 16 Sep 2026: “At full precision, a 30-billion parameter model would require over 55 GB of memory — far more than any consumer GPU offers.” The quantised model is “under 20 GB”, leaving room for “the perception encoder for image understanding, and the speculative decoding drafter to run simultaneously within a 24 GB or 32 GB envelope”.
TAKEAWAY

Running on a consumer GPU means the quantised build on a 24 GB card. A 30B model at full precision does not fit on anything you can buy for a desk.

How do you actually run it?

What Meta names: its status on 10 Aug 2026, updated where Meta or a partner has said more since (checked 22 Sep 2026)
RouteStatus
Hugging Face weightsavailable that day
Ollamashipped 10 Aug 2026
LM Studioshipped 10 Aug 2026 (Bionic)
Unslothguide and quants published
llama.cppb10353 or newer
ExecuTorchrepo listed
MLXvia Ollama on Apple Silicon
vLLM, SGLangfor serving at scale

Meta AI Research, Introducing Muse Glimmer, 10 Aug 2026: “In the coming days, run it locally through partners like Ollama, LM Studio, and Unsloth, deploy it with edge frameworks including llama.cpp, ExecuTorch, and MLX, serve it at scale with vLLM and SGLang”. Weights: meta-models/Muse-Glimmer-30B.

Meta, meta-models/Muse-Glimmer-30B-GGUF model card, read at source 22 Sep 2026: “You need llama.cpp build b10353 or newer.” and “Muse Glimmer support was merged on 10 Aug 2026”. The same card lists a related repo of “pre-exported ExecuTorch programs for NVIDIA CUDA and Apple Silicon”, not opened here. Until 22 Sep 2026 the llama.cpp and ExecuTorch rows read “in the coming days”.

Ollama, Muse Glimmer from Meta Superintelligence Labs is now available, 10 Aug 2026, read at source 22 Sep 2026: Muse Glimmer “is now available on Ollama”, run with ollama run muse-glimmer, and on Apple Silicon through Ollama’s MLX engine as muse-glimmer:30b-mlx. LM Studio, Run Muse Glimmer locally, 10 Aug 2026, read at source 22 Sep 2026: “We partnered with Meta to bring launch day support for Muse Glimmer in LM Studio Bionic!” Unsloth, Muse Glimmer - How to Run Locally, undated, read at source 22 Sep 2026: “You can run or fine-tune Muse Glimmer with Unsloth”; “Thank you Meta for providing Unsloth with day-zero support.”

First-hand: until 22 Sep 2026 the Ollama, LM Studio, Unsloth and MLX rows read “in the coming days”, and the text below said to “treat Ollama, LM Studio, Unsloth and MLX as announced, not confirmed.” Ollama and LM Studio had both shipped on launch day, five weeks before this page was written.

Most of those routes have shipped. Ollama and LM Studio both supported Muse Glimmer on launch day: on Ollama the command is ollama run muse-glimmer, and Macs with Apple Silicon can use Ollama’s MLX engine; in LM Studio it runs in the Bionic app. Unsloth publishes a guide and quantised files, and llama.cpp support is confirmed on Meta’s own model card, from build b10353. This page did not install any of them. MLX on its own, outside Ollama, is not confirmed here.Ollama, blog, and LM Studio, blog, both 10 Aug 2026; Unsloth, guide. All read at source 22 Sep 2026.

Why it is faster than its size suggests

Two tricks, both documented. Quantisation shrinks the weights, and a small companion model writes ahead of the big one. Meta: the drafter is “a small companion network that proposes entire blocks of tokens at once”, and “the main model then verifies these proposals in parallel, accepting correct tokens and correcting wrong ones.”

What is it good at?

Agent work, by design. Meta lists reliable tool use, multi-step reasoning and, unusually, failure recovery: “When a tool call fails or returns an unexpected result, the model is trained to diagnose the error and retry rather than halt.” It also takes images — “the model accepts interleaved text and images” — and Meta says it is “trained on data from more than 100 languages”.

The comparison Meta draws is against its own size class: it says Glimmer “performs strongly for its size class” next to Gemma4-31B and Qwen3.6-27B. That is a vendor benchmark claim, not an independent one.

What this page could not verify

  • Whether MLX works on its own. Ollama runs Muse Glimmer through its MLX engine on Apple Silicon; standalone MLX support, which Meta also announced, is not checked here. Ollama, LM Studio and Unsloth are confirmed by their own pages, llama.cpp by Meta’s GGUF card.
  • Real speed on any machine. Meta’s figures are Meta’s.
  • Benchmark scores. The announcement charts them without publishing the numbers in text.
SOURCES

Meta AI Research, Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device, 10 August 2026, read at source on 16 September 2026 · Meta, Muse-Glimmer-30B-GGUF model card, read at source on 22 September 2026 · Ollama, Muse Glimmer from Meta Superintelligence Labs is now available, and LM Studio, Run Muse Glimmer locally, both 10 August 2026 · Unsloth, Muse Glimmer - How to Run Locally. All three read at source on 22 September 2026.

Memory figures are the vendor’s, not measured here.

ABOUTMETHODVERIFYPRIVACYCONTACTINDEXAI PROMPT GENEER · EVERY ARTICLE CARRIES ITS OWN CHECKED DATE