Models

Muse Glimmer: Meta’s open model you can run yourself

Glimmer is the one Muse model you can download. Meta released it in August 2026 as an open-weight, 30-billion-parameter model built for agents that run on your own computer. It isn’t the Muse app, and it isn’t Muse Spark. Here’s what it is and what you need to run it.

By Published Facts checked 2026-10-06

Quick answer

Muse Glimmer is Meta’s open-weight AI model, released August 10, 2026: about 30 billion parameters, distilled from Muse Spark, under the permissive Apache 2.0 license. It’s built for AI agents that run locally: tool use, coding and multi-step tasks on “a Mac or PC with a single consumer GPU.” Download it from Hugging Face or run it with Ollama or LM Studio; the 4-bit builds need about 17–20 GB of memory. It’s a developer model, not the Muse app.

How we checked

Not testedWe haven’t run Glimmer ourselves. Details come from Meta’s announcement, Meta’s Hugging Face model card and Ollama’s listing, read October 6, 2026. Benchmarks are Meta’s. See how we label evidence.

What it is

Maker
Meta Superintelligence Labs, announced August 10, 2026.1
Size
About 30 billion parameters (the model card lists ~29.6B, dense, no mixture-of-experts), plus a vision encoder.2
Training
Distilled from Muse Spark, Meta’s larger model.1
License
Apache 2.0: commercial use, modification and redistribution allowed.1
Context
131,072+ tokens per the model card.2
Built for
“Always-on local agent workflows”: tool calling, coding, image understanding and using an AI model to grade other outputs.1

How it fits with Muse and Muse Spark

Meta’s “Muse” products compared
ProductWhat it isWhere it runs
MuseThe personal AI agent appMeta’s cloud (your private VM)
Muse SparkMeta’s main model family; powers Muse and Meta AIMeta’s cloud; Meta Model API
Muse GlimmerOpen-weight model for local agentsYour own computer, or partner clouds

What you need to run it

Per MetaMeta says Glimmer runs on “a Mac or PC with a single consumer GPU” and, at 4-bit, fits in 24–32 GB of memory.1 Meta’s Hugging Face model card lists two 4-bit files:2

  • Q4_K_M, 16.8 GB: fits 24 GB of VRAM.
  • Q4_K_XL (dynamic), 19.7 GB: for 32 GB.
  • Memory overall: about 17 GB for text only, about 19 GB with vision, about 20 GB with vision and the speed-up “drafter.”

Full-precision (BF16) weights are in a separate repository, meta-models/Muse-Glimmer-30B.2

Where to get it

Hugging Face
meta-models/Muse-Glimmer-30B (full) and meta-models/Muse-Glimmer-30B-GGUF (4-bit).2
Ollama
ollama run muse-glimmer:30b: an 18 GB build with vision, tools and thinking.3
Others
Per MetaLM Studio, Unsloth, vLLM, SGLang, and hosted on Together AI, Fireworks AI and OpenRouter.1

How good is it?

Per MetaMeta says Glimmer beats Gemma4-31B and Qwen3.6-27B on agentic, coding, multimodal, safety and reasoning benchmarks.1 The model card lists, among others: MCP Atlas 75.5, DeepSearch QA 74.6, SWE-Bench Pro 51.2, MMMU Pro 74 and AIME 2026 94.7%.2 These are Meta’s own results.

Safety notes

Per MetaMeta says Glimmer was evaluated under its Advanced AI Scaling Framework.1 The model card warns it “may produce inaccurate, biased, or objectionable responses” and isn’t optimized for video.2 Because it runs on your hardware, you’re responsible for how a local agent built on it is allowed to act.

Sources

  1. Introducing Muse Glimmer Meta AI Research · Aug 10, 2026 · read Oct 6, 2026
  2. Muse-Glimmer-30B-GGUF model card Meta on Hugging Face · read Oct 6, 2026
  3. muse-glimmer:30b Ollama library · read Oct 6, 2026

Last checked: 2026-10-06. New Glimmer versions may change these numbers; tell us if you spot one.

Frequently asked questions

Is Muse Glimmer open source?

Its weights are open under the Apache 2.0 license, which allows commercial use, changes and redistribution. Meta released it on August 10, 2026.1

What hardware do I need to run Muse Glimmer?

Meta says a Mac or PC with a single consumer GPU. The 4-bit builds on Hugging Face need roughly 17–20 GB of memory and are listed as fitting 24 GB (smaller build) or 32 GB (larger build) of VRAM.12

How do I run Muse Glimmer with Ollama?

Ollama lists it as muse-glimmer:30b (about 18 GB), run with ollama run muse-glimmer:30b. LM Studio, llama.cpp and vLLM are also supported.23

Does the Muse app use Glimmer?

Meta doesn’t say so. The Muse agent app runs on Muse Spark in Meta’s cloud. Glimmer is a separate model for developers who want an agent on their own hardware.1