Meta's Muse Models Run Where I Actually Need Them

Meta shipped Muse Spark 1.2 with a million token context, then open sourced Muse Glimmer, a 30B model that runs on one GPU. I ran it on my own machine and it changed how I think about local AI.

523 views2 min readPublished August 10, 2026

MetaOpen SourceLocal Models+1

Two Meta releases inside one week, and the second one is the reason I'm writing today.

Muse Spark 1.2 arrived August 5 as their paid agentic model, notable mostly for the million token context window and a terminal companion called Muse Code. Nice, but not what grabbed me.

Then Came Glimmer

On August 10 they released Muse Glimmer. Thirty billion parameters, multimodal, Apache 2.0 licensed, and sized to run on a single consumer GPU.

My homelab box has 16GB of VRAM : it fit
Downloaded through Ollama, working agent in ~15 minutes
Function calling worked without drama
Optimized builds landing for llama.cpp, MLX, ExecuTorch

First thing I pointed it at was summarizing my server logs, because that's my version of hello world.

It never once tried to hallucinate a log line that wasn't there, which is more than I can say for some bigger models I've tested on the same task. Meta also partnered with AMD, Arm, Dell, Intel and NVIDIA on device-level optimization, which tells you they expect this thing embedded everywhere.

Setup Notes if You Want to Try It

Grab the weights from Hugging Face or pull via Ollama directly
Leave headroom: quantized it idles around 9GB VRAM on my box
Start with the function-calling scaffold from their docs rather than rolling your own
Pair it with a small draft model if you care about tokens per second

Is It Frontier Smart? No. Does It Matter?

Obviously not as smart as Sol or Opus 5. But here's the thing: for maybe sixty percent of my daily tasks, it doesn't have to be.

Classification
Extraction
Quick summaries
Judging whether another model's answer is any good

That work is currently costing me API credits for zero good reason. There's also a privacy angle I didn't fully appreciate until the logs were running locally: nothing leaves the house. For a teenager whose browser history is ninety percent Stack Overflow and Minecraft wikis, that's mostly a curiosity. For anyone handling anything sensitive, it's the whole pitch.

Big closed models for the hard stuff, small open models living on your own hardware for everything else. My electricity bill says thanks either way.

Cheers, Yassen