New model Apache 2.0 August 10, 2026

Run Meta's Muse Glimmer 30B with Lemonade

Meta Superintelligence Labs released Muse Glimmer 30B today: a dense, multimodal, open-weights model trained for agentic work rather than one-shot answers. It reads screenshots and documents, calls tools, and manages its own memory, so a long multi-step task can pick up where it left off instead of starting over. Lemonade runs it locally in three commands, and serves it over the OpenAI-, Ollama-, and Anthropic-compatible APIs your agents already speak.

Parameters
30B dense
Input
Text and vision
License
Apache 2.0
Download at 4-bit
About 20 GB
Runs on
32 GB+ GPUs

Setup

Set aside about 20 GB of free disk space for the download.

1

Point llama.cpp at the latest release

Muse Glimmer is a new architecture, so it needs the upstream llama.cpp build that added support for it. Run the one that matches your hardware.

# vulkan: cross-platform, covers AMD, NVIDIA, Intel, and ARM GPUs lemonade config set llamacpp.vulkan_bin=latest
# rocm: optimized for AMD lemonade config set llamacpp.rocm_bin=latest
# cuda: optimized for NVIDIA lemonade config set llamacpp.cuda_bin=latest

Note: Set Lemonade to use any backend your hardware supports with the llamacpp config, for example lemonade config set llamacpp.backend=rocm.

Note: Set the same key to builtin to return to the llama.cpp version Lemonade ships and tests, for example lemonade config set llamacpp.vulkan_bin=builtin.

Note: Muse Glimmer ships as a built-in model in Lemonade v11.6.0, so this step becomes optional once you update.

2

Pull the weights

We like to recommend the 4-bit Unsloth Dynamic quantization for local use. Other sizes, from 2-bit up to BF16, are available in the same Unsloth repository. Lemonade picks up the vision projector from that repo automatically.

# pull: downloads about 20 GB, weights plus vision projector lemonade pull unsloth/Muse-Glimmer-30B-GGUF:UD-Q4_K_XL
3

Use the model

Lemonade registers the download as Muse-Glimmer-30B-GGUF-UD-Q4_K_XL, combining the repository name with the quantization you chose. Use that name with whichever of these fits where you want to work.

Serve the endpoint to apps and agents

lemonade load Muse-Glimmer-30B-GGUF-UD-Q4_K_XL

Open it in the web app

lemonade run Muse-Glimmer-30B-GGUF-UD-Q4_K_XL

Chat in your terminal

lemonade chat Muse-Glimmer-30B-GGUF-UD-Q4_K_XL

Use it in a coding agent

Claude Code, Codex, and OpenCode work here too.

lemonade launch pi --model Muse-Glimmer-30B-GGUF-UD-Q4_K_XL

Not running Lemonade yet? It installs in about a minute.

Install Lemonade