Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device

Meta Muse Glimmer 30B · open-weight · Apache 2.0 · on-device agents · multimodal
Interactive lab + deep-dive guide for the local agentic model released 10 August 2026

Weights · Research blog · HF day-0 guide · Article in this repo

---

Muse Glimmer is Meta Superintelligence Labs’ open-weight 30-billion-parameter agentic multimodal model optimized for always-on local agent workflows. It runs on consumer hardware (Mac / PC with a performant GPU or large unified memory) after 4-bit quantization (under 20 GB for the LM), with tool calling, failure recovery, controllable reasoning, and a 2B perception encoder for images and video frames.

This repository is a search-friendly companion: a long-form intro (BLOG.md) plus an offline-first interactive lab so you can feel agent loops, benchmarks, and memory envelopes before you download the weights.

Keywords: Muse Glimmer, Meta AI, open agentic model, local LLM, on-device AI, Apache 2.0, function calling, OpenClaw, llama.cpp, GGUF, multimodal agents, DFlash, SWE-Bench, MCP Atlas

---

The one-liner

An open agentic model that runs on your device — Apache 2.0 weights, dense 30B (2B perception + 28B decoder), ~4-bit under ~20 GB, trained for tool loops, failure recovery, and controllable reasoning.

Model	Muse Glimmer 30B (dense multimodal)
Weights	meta-models/Muse-Glimmer-30B
License	Apache 2.0
Released	10 August 2026
This repo	Offline-first lab + optional live OpenAI-compatible chat

---

30-second start

git clone https://github.com/cobusgreyling/Muse-Glimmer.git
cd Muse-Glimmer

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python app.py

Open http://127.0.0.1:7870

Or:

./run.sh

Docker

docker compose up --build
# → http://127.0.0.1:7870

No GPU. No API key. Scenario playback, benchmarks, and memory sizing work offline.

Optional live chat (llama.cpp / Hugging Face)

Full guide: docs/LIVE.md · Publish checklist: PUBLISH.md

A) On-device with llama.cpp

# Terminal 1 — serve GGUF (needs llama.cpp installed)
./scripts/serve-llamacpp.sh

# Terminal 2 — wire lab + run
./scripts/configure-live.sh llamacpp
./run.sh
# → Live chat tab → Probe endpoint → Send

B) Hugging Face Inference Endpoint

export HF_ENDPOINT_URL="https://XXXX.region.cloud.endpoints.huggingface.cloud"
export HF_TOKEN="hf_…"
./scripts/configure-live.sh hf-endpoint
./scripts/probe-live.sh   # lab must be running for full probe
./run.sh

Or hand-edit .env (see .env.example). Leave OPENAI_MODEL empty to auto-pick from GET /v1/models.

---

What the lab shows

Tab	What you learn
Agent loops	Multi-step local agents: tools, failure recovery, coding fix, multimodal tool call, reasoning low vs high
Benchmarks	Interactive scoreboard vs Gemma4-31B & Qwen3.6-27B (published launch numbers)
Memory	Footprint calculator — why ~4-bit + 24–32 GB is the practical envelope
Live chat	Optional /v1/chat/completions against your Muse Glimmer endpoint

Scenario highlights

1. Home Assistant dashboard — discover → 401 → recover token → build → deploy
2. Coding fix — red test → patch → green
3. Multimodal tools — image city → weather.get → clothing advice
4. Reasoning A/B — same architecture question, low vs high effort

---

Why Muse Glimmer matters

Most “local models” are chatty generalists squeezed onto a laptop. Muse Glimmer is aimed at a harder job:

• Always-on agents that own multi-step work on-device
• Privacy — contracts, code, home automation stay local
• Cost — no per-token meter for every tool hop
• Latency — no round-trip to a distant region for every thought

Meta’s claim (launch materials): strong agentic scores in the ~27–31B class, with deliberate training for function calling, long horizons, and diagnose-and-retry when tools fail — not just next-token fluency.

Details and nuance: BLOG.md.

---

Architecture (cheat sheet)

┌─────────────────────────────────────────────────────────┐
│  Muse Glimmer 30B (dense multimodal)                    │
│                                                         │
│   Perception Encoder (~2B ViT-style)                    │
│        │  images / video frames                         │
│        ▼                                                │
│   Text decoder (~28B)                                   │
│     · 52 layers · (SWA×3 + Full)×13                     │
│     · GQA (16 Q per KV) · Q-K norm                      │
│                                                         │
│   Optional: DFlash speculative drafter (block proposals)│
└─────────────────────────────────────────────────────────┘
         │
         ▼  ~4-bit quant  →  LM ≲ 20 GB
   24–32 GB device envelope (KV + vision + drafter headroom)

---

API

Method	Path	Description
GET	/	Lab UI
GET	/api/health	Demo / live status
GET	/api/model-card	Specs, hub links, capabilities
GET	/api/scenarios	Scenario list
GET	/api/scenarios/{id}	Full agent transcript fixture
GET	/api/benchmarks	Published comparison rows
GET	/api/prompts	Example prompts
POST	/api/footprint	Memory envelope estimate
GET	/api/live/probe	Reachability + models + tiny completion
GET	/api/live/models	List models from configured endpoint
POST	/api/chat	Live OpenAI-compatible chat (optional)

---

Project layout

Muse-Glimmer/
├── README.md
├── BLOG.md                 # Long-form intro article
├── PUBLISH.md              # Standalone repo publish checklist
├── docs/LIVE.md            # llama.cpp + HF live wiring
├── app.py                  # FastAPI lab
├── scripts/
│   ├── configure-live.sh   # .env presets (llamacpp | hf-endpoint | openrouter)
│   ├── serve-llamacpp.sh   # start llama OpenAI server
│   └── probe-live.sh       # health-check live endpoint
├── requirements.txt
├── run.sh
├── .env.example
├── assets/header.jpg
├── data/
│   ├── model-card.json
│   ├── agent-scenarios.json
│   └── prompts.json
└── static/                 # UI

---

Run Muse Glimmer itself (not just the lab)

Path	Link / note
Full weights	Hugging Face
GGUF	meta GGUF · Unsloth
Research blog	research.meta.ai
HF day-0 guide	huggingface.co/blog/muse-glimmer
transformers / llama.cpp / vLLM	Day-0 support called out in HF blog
Cloud	Together · Fireworks · OpenRouter · HF Inference Endpoints

# transformers sketch (needs sufficient VRAM / offload)
from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "meta-models/Muse-Glimmer-30B"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, dtype="auto", device_map="auto"
)

---

Stack

FastAPI · Uvicorn · vanilla HTML/CSS/JS · offline JSON fixtures · optional httpx → OpenAI-compatible API

---

Disclaimer

This lab is an independent educational companion, not an official Meta product.
Agent timeline transcripts are illustrative fixtures for teaching agent behavior.
Benchmark numbers are taken from public launch materials (Hugging Face / Meta, August 2026).
Model weights remain under Meta’s Apache 2.0 release; this repo’s code is MIT.

---

License

• Lab code & fixtures: MIT
• Muse Glimmer weights: Apache 2.0 (Meta)