Muse Glimmer: a 30B agentic model that runs on one GPU, no data center required
A serious agent, no data center required
Most "run it locally" model announcements come with an asterisk — smaller, weaker, a toy version of the real thing. Meta's newest release doesn't: Muse Glimmer, a 30B multimodal model built specifically for agentic work, fits on a single consumer GPU and beats larger models on the benchmarks that actually measure agent behavior.
Architecture
Muse Glimmer is 30B parameters total: a 2B ViT-style vision encoder bolted onto a 28B-parameter text decoder, 52 transformer layers using a hybrid attention pattern. It's distilled from Meta's larger Muse Spark model — the capability of a bigger model, compressed into something a single GPU can hold. Training data spans 100+ languages, with a January 4, 2026 knowledge cutoff.
Multimodal understanding
- Text, image, and video — video comprehension up to 96 frames at 2 fps (no audio track processed).
- Open-ended object detection — it can locate and identify objects in a scene without a predefined label set, rather than only recognizing a fixed category list.
- 131K+ token context window, long enough for extended agent sessions or large documents without external chunking.
Agentic tool use
The headline feature: multimodal tool-calling with structured outputs. It can look at an image and decide which function to call based on what it sees, not just parse text instructions — e.g. inspecting a screenshot and calling the right API based on what's rendered, not a text description of it. It also generates and executes code, and is built with explicit failure-recovery behavior rather than assuming every tool call succeeds on the first try.
Running it locally
- Fits on one consumer GPU — quantized variants run in 24–32GB of VRAM.
- Day-0 support in
transformers,llama.cpp,vLLM, and Inference Endpoints — no waiting on community ports. - DFlash speculative decoding — up to 3x faster generation on supported hardware.
- No cloud round-trip required for any of the above; everything runs on the box you own.
Licensing
Apache 2.0 — no commercial-use gate, no attribution requirement, no separate license negotiation to run it in a product. Worth stating plainly because it's not a given right now: some recent open-weight agentic models carry commercial-use restrictions that only surface once you read the license text closely. This one doesn't.
Meta also ran safety evaluations for chemical/biological, cybersecurity, and loss-of-control risk — all rated "moderate or lower."
How it compares
Meta's own published numbers, against Gemma4-31B and Qwen3.6-27B:
| Benchmark | Muse Glimmer | Gemma4 | Qwen3.6 |
|---|---|---|---|
| MCP Atlas (agentic) | 75.5 | 54.2 | 62.5 |
| DeepSearch QA | 74.6 | 61.7 | 71.1 |
| GAIA2 | 43.3 | 36.4 | 40.0 |
| SWE-Bench Pro | 51.2 | 36.9 | 50.2 |
| SWE-Bench Verified | 76.0 | 66.6 | 77.2 |
It's not a clean sweep — Qwen3.6 edges it on SWE-Bench Verified — but against Gemma4 the gap is wide, and against Qwen3.6 it's competitive or ahead on most agentic-specific benchmarks. (Numbers from Meta's Muse Glimmer announcement; benchmarks are directional, not gospel.)
Why it matters
The self-hosted AI story has always had a quiet tax: the good agentic models needed real infrastructure, so "run it yourself" often meant "run a worse version of it yourself." A model that's genuinely built agent-first, ships permissively licensed, and fits on hardware a single person can own is the gap closing in real time — the same trend this whole tutorial series has been betting on.
Is it on WEC?
Not yet — and we're not going to pretend otherwise. We're running it through our own model-evaluation suite now, the same one that benchmarks everything already on WEC Models, before it earns a place in the catalog. If it holds up against what's already there, expect it soon.
📖 Sources: Meta's Muse Glimmer announcement (Hugging Face) · Model card