Use Muse Glimmer locally with OpenClaw | HolaClaw Docs

Search Ctrl K Download 👋¡Hola!

Use Muse Glimmer locally with OpenClaw<br>Run Meta's Muse Glimmer 30B locally behind OpenClaw on your Mac: hardware requirements, llama.cpp and Ollama setup, and the one-click route with HolaClaw.<br>8 min read Updated August 12, 2026 TutorialsLocal modelsOpenClaw<br>Muse Glimmer is Meta Superintelligence Labs' first open model, released on August 10, 2026. It's a 30B multimodal model with a 128K+ context window, published under Apache 2.0 and built for always-on local agent workloads. That's perfect for OpenClaw, and running it locally means no API key, no per-token bill, and nothing leaving your Mac .

OpenClaw doesn't run models itself. It talks to an engine that serves one over a local endpoint, so to integrate Muse Glimmer you need to:

Serve the model on your Mac

Configure OpenClaw to use it

This guide covers different scenarios , from running your own inference server to using HolaClaw, our native macOS application that installs and runs OpenClaw securely on your Mac.

What you need

Before you start. On base M3 and M4 MacBook Airs, expect minutes of waiting before you get a response from OpenClaw. We recommend checking the Can your Mac run Muse Glimmer? article.

An Apple-silicon Mac. Muse Glimmer runs on the GPU through Metal, so M-series only.

32 GB of unified memory or more. Muse Glimmer, at the context length OpenClaw needs, does not fit in less.

About 30 GB of free disk. Around 17 GB of model weights, plus OpenClaw and the apps this guide installs.

M5 Pro or better chip. Base M3 and M4 chips land around 4.3 tokens per second; an M5 Pro or better is where an assistant starts to feel responsive.

OpenClaw installed and running. Only for options A and B below. If you don't have it yet, skip to If you are starting from zero.

Context length is the setting that decides whether the model fits. Bigger context means more memory, and an OpenClaw assistant needs a good amount of it: the system prompt, tool definitions, and skills already fill about 19K tokens before you type anything.

How to run OpenClaw with Muse Glimmer?

Serve the model with an inference engine on your Mac, then configure OpenClaw to use its endpoint. We prepared different guides based on your current setup. You can jump directly to the section that best covers your use case.

You already run OpenClaw. You need an engine serving Muse Glimmer. Use Ollama if you already have it, or Llama for the setup we benchmarked.

You're starting from zero or you are experimenting. Both OpenClaw and the model can be installed in a single step, either with ollama launch openclaw or with HolaClaw.

If you already run OpenClaw

Option A: serve it with Ollama

Install Ollama. Download the application from ollama.com and open it.

Pull the model. Muse Glimmer has been in the Ollama library since release day. Open a terminal and run:

Copyollama pull muse-glimmer:30b-mlx

Raise the context length. In the Ollama application, open Settings and move Context length to 64k . The default of 16k is below what an OpenClaw assistant needs before you even type.

Keep the Ollama application open. It serves the model to your OpenClaw installation.

Configure OpenClaw to use Ollama. Run the following command where your OpenClaw installation lives. If it's on the same Mac:

Copyopenclaw onboard --non-interactive --accept-risk --skip-health \<br>--auth-choice ollama \<br>--custom-base-url "http://localhost:11434" \<br>--custom-model-id "muse-glimmer:30b-mlx"

If OpenClaw runs on a different machine, you first need to expose the Ollama service to the network. For that, open the Ollama application, go to Settings , and turn on Expose Ollama to the network .

Then, run the following command in the machine that runs OpenClaw:

Copyopenclaw onboard --non-interactive --accept-risk --skip-health \<br>--auth-choice ollama \<br>--custom-base-url "http://OLLAMA_HOST:11434" \<br>--custom-model-id "muse-glimmer:30b-mlx"

You have all the documentation about the Ollama provider in the OpenClaw docs.

Check it works. Start a conversation and wait for the first reply. On base hardware it will take a while.

A note on engines: the numbers we publish come from llama.cpp with the GGUF build, which is what HolaClaw uses under the hood. Here we recommend using MLX instead, as it may give you some performance gains on Apple silicon.

Option B: serve it with Llama (llama.cpp)

Llama is a Mac native application that uses llama.cpp, an inference engine, under the hood. It's more complex than Ollama, but it gives you more configuration and options, and the new Llama application makes it much easier to use than it used to be.

Install Llama. Download the application from llama.app, or install it with brew install --cask llama-app.

Pull the model and serve it. You can use the llama CLI tool to install and serve the official Muse-Glimmer-30B-GGUF model. The kquant-17gb variant requires about 17 GB on disk. To install it, open a...