Muse Glimmer 30B

Muse Glimmer 30B

Reads text and images and carries out agent tasks such as tool calls and multi-step reasoning on a local machine, in a 30B model distilled from Muse Spark. For developers building local agents.

Made by meta-models on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Muse Glimmer is a model of about 30B parameters that reads text and images and answers in text. The card says it was distilled from Muse Spark and aimed at agent work on consumer hardware. It is meant to run on your own machine with no cloud service or network connection.

Who it suits: It suits developers who want a local agent that calls tools, reasons over several steps and reads screenshots.

Industry | Who uses it | How they use it | Example
Software | Agent builders | Run a local agent that calls tools | A coding agent that retries after a failed call
Operations | Analysts | Ask questions about charts and documents | A chat about a screenshot of a report
Research | ML researchers | Study a distilled agent model | A run on tool use and coding benchmarks

People without a machine that can hold a 30B model can skip it, since even the compressed builds target 24 GB of video memory or more.

Where to use it:

• Agent setups such as OpenClaw and Hermes Agent, which the card names
• Hugging Face Transformers, the library the card lists
• Windows, macOS or Linux, where you download the files with Python
• A local GPU, with 64 GB for full precision, or 32 GB or 24 GB for the compressed builds

Use case | Who | What they get back
Tool-using agents | Developers | Function calls with fitting schemas over long workflows
Screenshot and chart reading | Analysts | Answers about images, charts and documents next to chat
Multilingual chat | Researchers | Replies in over 100 languages the training data covers

How to set it up: You need these first:

• Python and the huggingface_hub package
• A GPU with enough memory, about 59.6 GB for BF16 weights or about 14.9 GB at 4-bit by estimate
• An agent setup or Transformers to run the model in

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Load the model locally and point your agent setup at it.

pip install -U huggingface_hub
hf download meta-models/Muse-Glimmer-30B

How to use it: Run the model locally and give it a task through your agent setup or a chat. You can attach images beside your text. The card says you can pick a reasoning effort to trade quality against speed.

Example: You show an agent a screenshot of a failing build and ask it to fix the problem. The card says the model can read the screenshot, call tools, and diagnose and retry when a call fails or returns something unexpected.

For newcomers and experts: An agent model does more than answer a question. It plans steps, uses tools and tries again after a mistake. A first small thing to try is a short chat with one image, to see how it replies before you connect any tools.

For experts, it is a dense transformer with a perception encoder of about 1.8B parameters. It has 52 layers where three local-attention layers alternate with one global layer, a sliding window of 2,048 and a context of 131,072 or more. The card says compressing the weights to about 4 bits brought a small average accuracy loss across 15 benchmarks. Quantizing trades a little accuracy for a smaller memory footprint.

You send text, with or without images
The perception encoder reads the images
The model plans steps and calls tools
A failed call is diagnosed and retried
You get a text answer

Good to know:

Title: Average accuracy loss after compression
K-Quant-Dynamic: 0.2%
K-Quant-17GB: 1.0%

• Size: about 30B parameters in Safetensors format, with up to 4,096 visual tokens per image
• Variants: Hugging Face lists 199 quantized and 67 fine-tuned versions made by other people
• Authors: the card names Meta Superintelligence Lab
• License: apache-2.0
• Cost to run: none named, no API key or account
• Data it reaches: it downloads model files from Hugging Face and works on the text and images you give it
• Not for: none named in the card

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 29.8B
• Weights: about 59.6 GB in BF16; about 14.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 165,489
• Likes: 2,143
• License: apache-2.0
• Quantized versions: 201
• Fine-tunes: 67
• Tags: transformers, safetensors, muse_glimmer, image-text-to-text, conversational

Links

• Model: https://huggingface.co/meta-models/Muse-Glimmer-30B
• Files: https://huggingface.co/meta-models/Muse-Glimmer-30B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download meta-models/Muse-Glimmer-30B

Then follow the usage in its model card. Nothing in this kit runs by itself.