Deploy Meta Muse Glimmer 30B Locally with Ollama

Meta released Muse Glimmer 30B on August 10, 2026 — a 30-billion-parameter open-weight model under Apache 2.0 that runs on a single consumer GPU. It is agentic (tool-use, agent workflows), multimodal (vision + text), and competitive with much larger models on coding and reasoning benchmarks. With Ollama, you can have it running locally in under 15 minutes.

• A local Muse Glimmer 30B server accessible at http://localhost:11434
• An OpenAI-compatible API endpoint for agents, coding tools, and test harnesses
• Vision capabilities — send images alongside text prompts
• No API keys, no cloud dependencies, no per-token costs

1. Pull the model

ollama pull muse-glimmer

Ollama will download the GGUF quantized version from its registry. This is approximately 18GB for the default Q4_K_M quantization. The download may take 10–30 minutes depending on your connection.

2. Run the model

ollama run muse-glimmer

This loads the model into VRAM and drops you into an interactive chat. Test it:

>>> What is 17 * 23?
389

>>> Describe what you see in this image: [paste image path]

3. Serve via the OpenAI-compatible API

Ollama automatically exposes an OpenAI-compatible endpoint. Point any OpenAI SDK client at it:

curl http://127.0.0.1:11434/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "muse-glimmer",
    "messages": [{"role": "user", "content": "Write a Python function to check if a number is prime."}]
  }'

4. Point your agent at it

In your agent's configuration, set the base URL:

base_url: http://localhost:11434/v1
model: muse-glimmer
api_key: ollama (any non-empty string works)

Hermes Agent, Continue, Open WebUI, and any OpenAI-compatible client will work without code changes.

5. Test multimodal (vision)

curl http://127.0.0.1:11434/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "muse-glimmer",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "What is in this image?"},
        {"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}}
      ]
    }]
  }'

• ollama ps shows muse-glimmer loaded with memory usage
• The curl chat completion returns a valid response
• ollama show muse-glimmer displays model metadata and parameters
• An agent pointed at localhost:11434/v1 completes a task successfully

• OOM on load: You need at least 24GB VRAM for Q4. If you have 16GB, try a smaller quantization or use a smaller model. Ollama does not let you pick quantization from the CLI — you would need to import a specific GGUF file for a lower quant.
• Slow first response: Cold load is expected. The first prompt loads weights into VRAM. Subsequent prompts are fast. Ollama unloads models after idle timeout (configurable via OLLAMA_KEEP_ALIVE).
• Vision not working: Ensure you are using the latest Ollama version that supports multimodal models. Check ollama --version is recent.
• Out of memory with other GPU workloads: Ollama keeps the model in VRAM until idle timeout. If you are running other GPU tasks, reduce the timeout: OLLAMA_KEEP_ALIVE=5m ollama run muse-glimmer.

• Throughput is for development, not production serving. Ollama is optimized for developer convenience, not high concurrency. If you need to serve many concurrent requests, use vLLM with Muse Glimmer instead. Ollama is for development and testing; vLLM is for production.
• Q4 quantization loses some quality. The full-precision model scores higher on benchmarks. For development and testing, Q4 is fine. For production quality, consider Q8 or FP16 if your VRAM allows.
• Muse Glimmer is new. The model shipped August 10, 2026. Ollama registry support and community quantizations are still maturing. If ollama pull muse-glimmer fails, check HuggingFace for GGUF files and import manually: ollama create muse-glimmer -f Modelfile.