Muse Glimmer 30B

Meta (Superintelligence Labs) · Released: 2026-08 (weights on Hugging Face Aug 9; Meta paper published Oct 19) · Last verified 2026-10-19

Start here Easy: Ollama

Easiest: Ollama, one command

Install Ollama (free) from ollama.com/download and open it once.
Open Terminal (Mac/Linux) or PowerShell (Windows) and paste the command below. The first run downloads the model.
Type your question and press Enter. Type /bye to quit.

ollama run muse-glimmer

Size: 18 GB download
You need: A 24 GB graphics card or a Mac with 32 GB of memory.
Good to know: It reads images too: drag a picture into the Ollama app, or put the image's file path in your message in Terminal. For harder coding or agent tasks, tell it how hard to think: /set system "Reasoning strength: high".

On a Mac, the MLX build (19 GB) has open bug reports (slow replies, stuck on Stopping), so start with the command above. Use this only if you want to try it:

ollama run muse-glimmer:30b-mlx

Checked 2026-10-19 · source · Everything below is the detail for people tuning their setup.

The details

Meta's 30B open model for chat, images and agents (Apache 2.0). One Ollama command, about 18 GB, fits a 24 GB graphics card or a 32 GB Mac.

Recommended quants

Official GGUF KQuant-17GB Q4_K_M (16.8 GB file)
meta-models/Muse-Glimmer-30B-GGUF
Meta's 'start here' build for 24 GB VRAM (~1.0% average degradation per Meta). Add mmproj (1.4 GB) for images and the optional DFlash drafter (1.6 GB) for faster replies.
Official GGUF KQuant-Dynamic Q4_K_XL (19.7 GB file)
meta-models/Muse-Glimmer-30B-GGUF
Higher-quality build aimed at 32 GB VRAM (~0.2% average degradation per Meta).
Unsloth Dynamic GGUF (UD-Q2_K_XL … UD-Q8_K_XL)
unsloth/Muse-Glimmer-30B-GGUF
Unsloth guide: 2-bit 12-14 GB, 4-bit 17 GB+, 6-bit 20-22 GB, 8-bit 34 GB total memory. Full BF16 is about 58 GB.

Chat template Works

Official chat_template.jinja ships with the weights. Reasoning strength (low / medium / high / xhigh) is set with a line in the system prompt, not a flag.

Fix: Add the system line "Reasoning strength: high" for coding and agent work, lower for quick chat. In the Ollama chat window: /set system "Reasoning strength: medium".

Sampling settings

Official	`temp 1.0, top_p 0.95, top_k 64`
Reasoning strength	`low / medium / high / xhigh via "Reasoning strength: <value>" in the system prompt`
Context	`131,072 tokens (Unsloth: up to 262,144)`

Runtimes

Ollama	Works	Official library muse-glimmer: :30b / latest (18 GB, text + image) and :30b-mlx (19 GB) for Apple Silicon, plus DFlash tags, as of Oct 19.
llama.cpp	Works	Needs build b10353 or newer (support merged Aug 10). Older builds refuse to load it; if replies start with to=self<\|message\|>, update.
LM Studio	Works	Listed as meta/muse-glimmer; LM Studio says the smallest build needs at least 26 GB of RAM.
MLX	Works	Via Ollama's MLX tags (:30b-mlx), which Ollama says support DFlash and image input.
vLLM	Unverified	Not checked for this card.

Tool calling in agent tools

Claude Code (local endpoint)	Unverified	Ollama documents `ollama launch claude --model muse-glimmer`.
OpenCode	Unverified	Ollama documents `ollama launch opencode --model muse-glimmer`.

Known issues

Sources checked

Spotted something wrong or a new fix? Email corrections to [email protected]. Corrections are reviewed before publishing.