Muse Glimmer + OpenClaw: Run Meta's Agent Model Offline

Bottom Line

• Muse Glimmer is the Meta model to use for a local agent. It is a 30B model under Apache 2.0. Meta’s model card names OpenClaw and Hermes Agent.
• Three commands: ollama pull muse-glimmer:30b, openclaw onboard (choose Ollama, Local only), openclaw models set ollama/muse-glimmer:30b.
• Raise the context. OpenClaw caps local Ollama models at 32,768 tokens by default. Glimmer’s cache is about 1.7 GiB at 131K, so set num_ctx to 131072.
• On a Mac, use the 30b-mlx tag. It is 19GB and includes the DFlash drafter.
• Muse Code cannot run locally. OpenClaw’s Muse plugin accepts only https://api.meta.ai/v1.
• Hardware: a 24GB GPU or a 32GB Mac. Our Glimmer VRAM page has every quant.

Need a machine for it? See the tested hardware list with current prices.

Which Meta Model or Agent Can Run Locally?

Meta ships no local agent framework of its own. The Hugging Face launch blog points to OpenClaw, Hermes and Pi. So the local Meta agent is Glimmer inside one of those runtimes.

Option 1: Ollama (Easiest)

Step 1: Pull the model

On an NVIDIA or AMD GPU:

ollama pull muse-glimmer:30b

On Apple silicon, use the MLX tag. It includes DFlash:

ollama pull muse-glimmer:30b-mlx

Both tags support images, tools and thinking. Test it:

ollama run muse-glimmer:30b "Reply with: ready"

Step 2: Connect OpenClaw

openclaw onboard

Choose Ollama, then Local only. If OpenClaw does not find Ollama, opt in to local detection. Any value works for a local host:

export OLLAMA_API_KEY="ollama-local"

Step 3: Set Glimmer as the primary model

openclaw models set ollama/muse-glimmer:30b

Use ollama/muse-glimmer:30b-mlx on a Mac. Check it:

openclaw models list --provider ollama

Do not point OpenClaw at Ollama’s /v1 URL. The OpenClaw docs say the OpenAI-compatible URL breaks tool calling. The model then prints raw tool-call JSON as text. Use the native URL, http://127.0.0.1:11434, which is what onboarding sets.

Raise the Context Above 32K

OpenClaw caps local Ollama models at 32,768 tokens unless you set params.num_ctx. An agent that reads files and tool output fills 32K fast.

Glimmer keeps the full context in only 13 of its 52 layers, so a long context is cheap:

Add this to your OpenClaw config. It follows the per-model params format in the OpenClaw Ollama docs:

{
  agents: {
    defaults: {
      models: {
        "ollama/muse-glimmer:30b": {
          params: { num_ctx: 131072 },
        },
      },
    },
  },
}

On a 24GB card, Q4 weights plus a 131K cache is about 18 GiB. That leaves room. On a 16GB card, stay at 32K.

Set the Reasoning Level

Meta’s card lists four levels: low, medium, high and xhigh. It suggests high or xhigh for coding and complex agent work. Meta sets the level in the system prompt with this line:

Reasoning strength: high

OpenClaw also has its own /think command for models that report thinking support. Use one method, not both, so you know which one is active.

Meta’s sampling settings are temperature 1.0, top_p 0.95 and top_k 64.

Option 2: llama.cpp With DFlash (Fastest on NVIDIA)

Meta measured 233.4 tok/s on an RTX 5090 with DFlash, against 74.9 without. The Hugging Face blog runs it like this:

llama serve -hf meta-models/Muse-Glimmer-30B-GGUF --spec-type draft-dflash --spec-draft-n-max 15

The server and its web UI run on localhost:8080.

Connect OpenClaw with the bundled llama.cpp provider:

openclaw plugins install @openclaw/llama-cpp-provider

It can use an existing llama-server. Model references use the form llama-cpp/<model>. See the OpenClaw llama.cpp plugin docs for discovery.

Test the Agent

Restart the gateway, then send one message. These commands are from the Hugging Face blog:

openclaw gateway restart
openclaw agent --message "Reply with: muse-ready"

If the reply is muse-ready, disconnect from the internet and send a real task. The model runs on your machine, so the agent keeps working.

Why Muse Code Cannot Run Locally

Muse Code is Meta’s terminal coding agent. OpenClaw can run it through the @openclaw/muse plugin, but the plugin has hard limits:

• Meta’s endpoint only. It accepts only the meta provider on https://api.meta.ai/v1. Any other baseUrl is refused, because Muse Code calls Meta directly.
• A Meta credential is required. Sign in, muse login, or META_API_KEY. Without one, the turn fails.
• No OpenClaw sandboxes or tools. Muse uses its own shell and file tools and its own MCP configuration.

So “Meta agent, offline” means Glimmer in OpenClaw’s own runtime. If you want Muse Code itself, you pay for Meta’s API and your code leaves the machine.

Honest Caveats

1. We did not run this setup. The commands come from the OpenClaw docs and the Hugging Face blog, checked on 31 October 2026.
2. The Ollama tag names may change. Check ollama list after the pull.
3. The cache figure is our estimate from config.json.
4. DFlash speeds are Meta’s. We have not seen independent numbers.
5. Ollama lists a 128K context, config.json says 131,072. Use 131072 for num_ctx.

Sources

• Hugging Face blog: Muse Glimmer: agent frameworks, llama.cpp and DFlash commands, OpenClaw test commands
• Muse Glimmer 30B model card: OpenClaw and Hermes, reasoning levels, sampling, DFlash speeds
• Ollama library: muse-glimmer: tags and sizes
• OpenClaw: Ollama setup: onboarding, OLLAMA_API_KEY, models set, verification
• OpenClaw: Ollama advanced configuration: 32,768 cap, params.num_ctx, thinking control
• OpenClaw: Ollama provider: native API, /v1 warning
• OpenClaw: custom providers: llama.cpp plugin install and refs
• OpenClaw: Muse plugin: Meta endpoint only, credentials, limits

See Also

• Can I run Muse Glimmer 30B locally?: VRAM by quant
• Qwen3.8-27B vs Muse Glimmer 30B: the other 24GB agent model
• How to run OpenClaw free and offline with Ollama: the general offline setup
• OpenClaw defaults to Claude, not Ollama: if the primary model will not change
• Best local models for OpenClaw: other models that call tools well