Muse Glimmer on an M4 Pro MacBook

A reproducible record of downloading and running Meta's Muse Glimmer 30B
model locally with llama.cpp and Apple Metal.

Tested successfully on August 10, 2026:

• 16-inch MacBook Pro
• Apple M4 Pro
• 24 GB unified memory
• muse-glimmer-30B-kquant-17gb.gguf
• llama.cpp commit 4801e3c

This setup is currently text-only. Image input requires the separate
mmproj-kquant.gguf perception encoder.

What is tracked here

Path	Purpose
README.md	Reproducible quick start
notes/setup.md	What we did, including mistakes and fixes
notes/how-local-inference-works.md	How a 17 GB model runs from disk and unified memory

Model weights, the Python environment, caches, and the local llama.cpp
checkout are intentionally ignored by Git.

1. Create the project and optional download environment

The Python virtual environment is only needed for the Hugging Face download.
It is not used when the model runs.

mkdir -p muse-glimmer-project
cd muse-glimmer-project

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade huggingface_hub

2. Download the model from Hugging Face

hf download meta-models/Muse-Glimmer-30B-GGUF \
  muse-glimmer-30B-kquant-17gb.gguf \
  --local-dir .

The download is approximately 16.8 GB. It is normal for ls -lh to display
this as roughly 16 GB because tools can use different unit conventions.

The environment can now be deactivated:

deactivate

3. Get and build llama.cpp

Muse Glimmer uses a new architecture, so use a current llama.cpp checkout.

brew install cmake
git clone --depth 1 https://github.com/ggml-org/llama.cpp.git llama.cpp
cd llama.cpp

cmake -B build \
  -DBUILD_SHARED_LIBS=OFF \
  -DGGML_METAL=ON \
  -DLLAMA_CURL=OFF \
  -DCMAKE_BUILD_TYPE=Release

cmake --build build --config Release -j 8 \
  --target llama-cli llama-server

A successful build ends with:

[100%] Built target llama-cli
[100%] Built target llama-server

4. Verify the binary

./build/bin/llama-cli --version

The tested shallow checkout reported:

version: 1 (4801e3c)
built with AppleClang 21.0.0.21000101 for Darwin arm64

version: 1 is expected after a --depth 1 clone: Git can see only one local
commit. The commit hash and Muse Glimmer architecture support ar