Muse Glimmer review
Meta's free 30B model that runs on one gaming GPU or a high-end Mac, made for local assistants.
Meta's free model that runs on one gaming GPU or a high-end Mac, made for local assistants that use tools. Strong at tool use and math for its size. Qwen3.6-27B is better at coding and computer use.
What is Muse Glimmer?
Muse Glimmer is a free AI model from Meta, released on 10 August 2026 under the Apache 2.0 license. It has 30 billion parameters, understands text and images, and is built to run on your own device as an always-on assistant that can call tools, write code and organise files. It is the closest thing to a new Llama this year.
Key features
- Free for any use, including business (Apache 2.0)
- Runs on one 24 GB graphics card or an M4/M5 Max Mac in its compressed 17 GB version
- Understands images as well as text, trained on 100+ languages
- Strong tool use: 75.5 on MCP Atlas
- Up to 3x faster with its built-in draft helper
Specs and price
| Made by | Meta |
| Released | 10 Aug 2026 |
| Size | 30B parameters |
| License | Apache 2.0: free, business use allowed |
| Input | Text and images |
| How much it can read | About 131,000 tokens |
| Computer needed | 24 GB GPU (17 GB version), 32 GB for the higher-quality version, or an M4/M5 Max Mac |
| Speed (RTX 5090) | Up to 233 tokens a second with the draft helper |
| Download | Hugging Face, meta-models/Muse-Glimmer-30B |
| Price | Free |
Test scores
| Test | Score | Published by |
|---|---|---|
| MCP AtlasUsing tools | 75.5 (Qwen3.6-27B: 62.5, Gemma4-31B: 54.2) | Meta |
| AIME 2026Math competition | 94.7 | Meta |
| SWE-Bench ProFixing real code bugs | 51.2 | Meta |
| IFBenchFollowing instructions | 77.0 | Meta |
| OSWorld-VerifiedUsing a computer | 65.9 (Qwen3.6-27B: 75.6) | Meta |
Scores published by Meta are the maker's own numbers. We will add our own results in the hands-on review.
How to download and run Muse Glimmer
- Download it from Hugging Face (
meta-models/Muse-Glimmer-30B). Pick the 17 GB version for a 24 GB GPU, or the higher-quality version for 32 GB. - Run it with Ollama, LM Studio, llama.cpp, or MLX on a Mac. Meta named these as launch partners.
- No GPU? Use it online through Together AI, Fireworks AI or OpenRouter.
- Add safety checks before letting it act on its own. Meta's tests found hidden bad instructions fooled it 28.4% of the time.
Who it is for
Good fit
- People who want a private assistant that runs offline
- Developers building local agents with tool calls
- Mac users with an M4 or M5 Max
Skip it if
- Computer-use and coding agents: Qwen3.6-27B scores higher
- Anyone who needs audio input
Pros and cons
Pros
- Free, with a business-friendly license
- Runs on a single consumer GPU or a Mac
- Best-in-class tool use for its size
- Fast with the draft helper
Cons
- Behind Qwen3.6-27B on computer use and SWE-bench Verified
- Compressed versions lose a little quality
- No audio, and video only as single frames
- Needs extra safety checks for unattended agents
How it compares
Against Qwen3.6-27B and Gemma4-31B, Meta's numbers show Muse Glimmer ahead on tool use, research and math. Qwen3.6-27B leads on computer use and SWE-bench Verified (77.2).
Hands-on review In testing
We are running Muse Glimmer on real tasks. Results will be added to this page, at the same address.
How it got here
| Date | What happened | Source type |
|---|---|---|
| Aug 10, 2026 | Muse Glimmer released under Apache 2.0.Meta, Introducing Muse Glimmer (10 Aug 2026) | Confirmed |
Frequently asked questions
Is Muse Glimmer the same as Llama 5?
No. It is part of Meta's newer Muse family. Meta has released no new Llama since April 2025, so Muse Glimmer is Meta's free model for 2026.
Is Muse Glimmer free?
Yes. It is released under Apache 2.0, so you can download it and use it for free, including in business.
Can Muse Glimmer run on my Mac?
Yes, on an M4 Max or M5 Max Mac with enough memory, using the compressed version. On a PC it needs a 24 GB graphics card.
Where can I download Muse Glimmer?
From Hugging Face at meta-models/Muse-Glimmer-30B, or through Ollama and LM Studio.
Related models
Muse Spark (free download)
MetaMeta's free version of Muse Spark is out: Muse Spark Open, 816B parameters with about 42B at work per word, on Hugging Face since 15 September and on OpenRouter since 15 October.
Llama 5
MetaStill searched a lot, but Meta has released no new Llama model since April 2025.
Flint
RecursiveFlint is out. Launch facts from official sources are below, and our hands-on review follows.