Get the model

Download Muse Glimmer from Hugging Face and pick the build that matches your runtime. Everything runs offline once the weights land: there's no API key and no gated access.

Prerequisites

• Python 3.10 or later
• The Hugging Face CLI: pip install -U huggingface_hub

Choose an artifact

Four repositories cover the published builds. Which one you want follows the runtime you plan to serve with:

The files inside the GGUF repository:

Download

The bf16 checkpoint, for vLLM and SGLang:

bash
hf download meta-models/Muse-Glimmer-30B --local-dir ./muse-glimmer-30b

The GGUF builds, fetching only the files you need:

bash
hf download meta-models/Muse-Glimmer-30B-GGUF --local-dir ./muse-glimmer \
  --include "Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf" \
  --include "mmproj-Muse-Glimmer-30B-Q4_K_M.gguf"

Add --include "dflash-Muse-Glimmer-30B-Q4_K_M.gguf" for the GGUF draft model.

The DFlash draft checkpoint:

bash
hf download meta-models/Muse-Glimmer-30B-assistant --local-dir ./muse-glimmer-assistant

Check the tokenizer

The main repository ships the tokenizer, chat_template.jinja, config.json, and processor_config.json, so it carries everything a runtime needs to frame prompts. Load the tokenizer and render a prompt to confirm the download is usable:

Python
from transformers import AutoTokenizer
​
tok = AutoTokenizer.from_pretrained("meta-models/Muse-Glimmer-30B")
​
messages = [{"role": "user", "content": "In one sentence, what is Muse Glimmer good at?"}]
print(tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False))

The rendered prompt uses Muse Glimmer's channel-scoped framing (<|start|>, <|message|>, <|eot|>), which the prompting guide covers.

Next steps

With the weights downloaded, set up your prompting format to get the best results from Muse Glimmer. When you're ready to serve, pick a runtime: vLLM, SGLang, llama.cpp, or ExecuTorch.