Muse Glimmer 30B GGUF by meta-models

Muse Glimmer 30B GGUF by meta-models

GGUF builds of a 30-billion-parameter agentic model for llama.cpp, with an optional image encoder and a speculative decoding drafter. They suit people who run models locally on a 24 GB GPU.

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: This repository holds GGUF conversions of Muse Glimmer 30B for llama.cpp. The card calls the model a 30-billion-parameter language model with its own perception encoder, built for agent tasks on consumer hardware and meant to run locally with no cloud access. You get two quantized text builds, an image encoder file and an optional drafter file for speculative decoding. The card names Meta Superintelligence Lab as the authors.

Who it suits: It suits people who want a large agent-style model running on their own machine through llama.cpp. A 24 GB GPU handles the smaller text build.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Local AI | Hobbyists with a 24 GB GPU | Run a private assistant with no cloud calls | Serve the model on the same computer and chat with it
Software | Agent developers | Point tool-using agents at a local server | Test multi-step tasks against a model you control
Research | Multimodal tinkerers | Add image input with the perception encoder | Load the encoder file next to the text build

If you need full precision, you can skip it. No bf16 GGUF is published here, and the card sends you to the base repository for that.

Where to use it: It runs in llama.cpp, using llama-server, llama-cli or llama-mtmd-cli. The card needs a build numbered b10353 or newer.

• llama.cpp on Windows, macOS or Linux
• A CUDA GPU, a CPU only, or Apple Metal, depending on how you build
• An OpenAI-style chat endpoint served by llama-server

Title: Jobs it does
Use case | Who | What they get back
Local chat and reasoning | Anyone with a big GPU | Answers, with the thinking returned in a separate field
Image questions | Multimodal users | Text answers about a picture, once the encoder file is loaded
Faster generation | Power users | Speculative decoding with the drafter file

How to set it up: Get a recent llama.cpp first, because older builds cannot load these files at all. Then download the weights you need and start the server.

• llama.cpp release b10353 or newer
• About 17 GB of memory for the smaller text build, more with extras
• Python with pip, for the Hugging Face tools

1. Download a llama.cpp release tagged b10353 or later, or build it from source.
2. Check the version number is at least 10353.
3. Download the text build and the encoder file.
4. Start llama-server with the model and the encoder.

./llama-cli --version     # build number must be >= 10353

cd llama.cpp
cmake -B build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build build --config Release -j \
    --target llama-cli llama-mtmd-cli llama-server

pip install huggingface_hub
hf download meta-models/Muse-Glimmer-30B-GGUF \
    --local-dir Muse-Glimmer-30B-GGUF \

Put your own value in its key setting first.

How to use it: Run llama-server with the text build and the encoder file, then send chat requests. The answer arrives in the content field and the thinking in a separate reasoning field. To add the drafter, pass its file with the extra flags.

./build/bin/llama-server \
    -m       Muse-Glimmer-30B-GGUF/Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf \
    --mmproj Muse-Glimmer-30B-GGUF/mmproj-Muse-Glimmer-30B-Q4_K_M.gguf \
    -a muse-glimmer-30B \
    -ngl 99 -c 131072 -np 4 \
    --host 127.0.0.1 --port 8080 \
    --jinja \
    --temp 1.0 --top-p 0.95 --top-k 64

    -md Muse-Glimmer-30B-GGUF/dflash-Muse-Glimmer-30B-Q4_K_M.gguf -ngld 99

Title: The files
Tool | What it does
Q4_K_M text build | The main text model, the card's starting point for 24 GB cards
Dynamic Q4_K_XL build | A higher-quality text build aimed at 32 GB cards
Perception encoder | Required for image input
DFlash drafter | Optional, for speculative decoding

Example: You start llama-server with the smaller text build and the encoder. You ask a short arithmetic question. The reply holds the answer in the content field, and the model's reasoning shows up separately.

For newcomers and experts: If you are new, this is a big AI model packed into files that llama.cpp can run on your own computer. Check your llama.cpp version first, then try the smaller text build alone.

If you know the field, these are 4-bit K-quant builds of a 30 billion parameter model with a separate multimodal projector and a draft model for speculative decoding. The text builds work alone, and the two companion files are additive. Only quantized files are published.

You install llama.cpp b10353 or newer
You download the text build and encoder
llama-server loads them on your GPU
You send chat requests to the server

Good to know: Memory use grows as you add the encoder and the drafter. Totals below include a working context.

Title: File sizes
17 GB text build: 16.8 GB
Dynamic text build: 19.7 GB
Perception encoder: 1.4 GB
DFlash drafter: 1.6 GB

Title: Rough total memory
Build | Text only | With vision | With vision and drafter
Smaller build | about 17 GB | about 19 GB | about 20 GB
Dynamic build | about 20 GB | about 22 GB | about 23 GB

• License: Apache 2.0
• Cost to run: none named, though you need a GPU with about 24 GB for the smaller build
• Data it reaches: the model files from Hugging Face, and the chats you send to your local server
• Not for: full precision, or llama.cpp builds numbered b10344 and older

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 27.9B
• Weights: about 13.9 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 198,352
• Likes: 393
• License: apache-2.0
• Based on: meta-models/Muse-Glimmer-30B
• Fine-tunes: 2
• Tags: gguf, image-text-to-text, conversational

Links

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

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