Muse Glimmer 30B GGUF

Muse Glimmer 30B GGUF

Runs a 30B agent model locally that reads text and images, calls tools and retries failed calls, from Unsloth's GGUF files. For developers building local agents on consumer hardware.

Made by unsloth 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 30B dense language model with a dedicated perception encoder, built for agent work on consumer hardware. The card says it was distilled from Muse Spark and runs locally with no cloud or network access. This repository holds Unsloth's GGUF files for it. The model takes text and images in and writes text out.

Who it suits: It suits developers who build agents that run on their own machines, and people who want an assistant that can look at screenshots, charts and documents.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Run a local agent that calls tools and recovers from failures | Retry after a tool call returns an error
Productivity | Power users | Ask about screenshots, charts and documents | Explain a chart from a screenshot
Research | ML researchers | Study a distilled agent model and its quantizations | Compare Unsloth's quantized files

Anyone without about 14 GB of room for the 4-bit weights can skip it.

Where to use it:

• Unsloth, where the card says it runs with toggles for thinking
• Agent scaffolds such as OpenClaw and Hermes Agent, and other orchestration patterns
• Your own machine, with no cloud service
• Windows, macOS or Linux, where the files download with Python

Use case | Who | What they get back
Finish a multi-step task | Agent developers | A completed workflow built from several tool calls
Write and debug code | Developers | Code and fixes inside an agent scaffold
Read a screenshot or chart | Everyday users | A text explanation of what the image shows

How to set it up:

• Python with pip, for the download
• A runtime that can load GGUF files, such as Unsloth
• Memory for about 13.9 GB of weights at 4-bit, as an estimate

1. Install the Hugging Face hub client.
2. Download the repository's files.
3. Load a quantization in Unsloth or another GGUF runtime.

pip install -U huggingface_hub
hf download unsloth/Muse-Glimmer-30B-GGUF

How to use it: Run the model in Unsloth and switch thinking on or off with its toggles. The card points to Unsloth's own guide for quantization analysis and step-by-step instructions.

Tool | What it does
Unsloth | Runs the model locally, with toggles for thinking
Perception encoder | Lets the model read text and images together, with up to 4,096 visual tokens per image
Effort control | Picks a balance between answer quality and speed
Tool calling | Calls functions with exact schemas across long workflows

Example: You give it a screenshot of an error message and ask it to find the cause. If a tool call fails along the way, the card says the model works out why and tries again rather than stopping.

For newcomers and experts: Think of it as a local assistant that can look at pictures and use tools. Start with a plain text question in Unsloth, then try a screenshot.

The language model is a dense causal transformer with 52 layers, a hidden width of 6656 and grouped-query attention with 32 query and 2 key-value heads. Attention repeats three local layers, then one global layer, with a 2048-token sliding window. RoPE applies to the local layers only. The perception encoder is a ViT-G/14 of about 1.8B parameters with 50 layers.

Text and images go in
The perception encoder reads the images
The 30B model plans and calls tools
A failed call is diagnosed and retried
Text comes out

Good to know:

• Context: 131,072 tokens or more
• Languages: trained on data from more than 100
• Size: the card says about 29.6B parameters with the vision encoder, and the record lists 27.9B for these files
• Tool calls: Unsloth's page says its 2-bit version ran more than 100 tool calls in Unsloth
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the text and images you give it, and the card says it needs no network access
• Not for: nothing named in the card

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.9B
• Weights: about 13.9 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 692,334
• Likes: 592
• License: apache-2.0
• Based on: meta-models/Muse-Glimmer-30B
• Quantized versions: 2
• Tags: transformers, gguf, unsloth, meta, image-text-to-text, conversational

Links

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

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download unsloth/Muse-Glimmer-30B-GGUF

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