How to run a free Meta model on an office Mac in about two hours

What Meta shipped

Meta released Muse Spark Open with open weights and an Apple partnership behind it: about 816B total parameters, about 42B active per token, context up to 1M tokens, and text, image, audio and video as inputs (Hugging Face).

Memory is the catch. Every expert stays resident, so the machine needs room for the whole 816B model while computing at the pace of a 42B one. Hugging Face ran the 3-bit MLX quant on an M3 Ultra with 512 GB of unified memory: about 344 GB resident, about 21 tokens a second generated. Even 3-bit needs more than 256 GB.

Meta's own setup is two Mac Studios with M5 Ultra and 512 GB each, treated as one machine (Meta). Meta's AI chief said most personal computers are too small to run powerful open-weight models (CNBC).

So "runs on a Mac" is accurate and, for an office, beside the point.

The model you can actually install

Muse Glimmer is the small dense model in the same series, released in August under Apache 2.0, which permits commercial use with no vendor agreement (VentureBeat). It is 30B parameters, text and images in, same tokenizer as the big one.

Unsloth's hardware table is the part to read before you download anything:

Published scores put Glimmer at 76.0 on SWE-Bench Verified and 75.5 on MCP Atlas (Hugging Face). Benchmarks are not your inbox, but this is not a toy either.

About two hours, start to first real task

These are my estimates from the documentation, not a measured run on your machine.

1. Check free unified memory, pick a build from the table: about 10 minutes.
2. Install the runner: about 25 minutes. Either the one line installer, or build llama.cpp from source (on a Mac set CUDA off, Metal is on by default).
3. Download the quant plus the vision projector file: about 40 minutes on a decent connection.
4. First run with Meta's sampler defaults: about 15 minutes.
5. Point one recurring job at it, classify yesterday's inbound or draft one reply, and log what it gets wrong: about 30 minutes.

curl -fsSL https://unsloth.ai/install.sh | sh
unsloth run --model unsloth/Muse-Glimmer-30B-GGUF:UD-Q4_K_XL

Or the llama.cpp path, with the settings Meta recommends:

./llama.cpp/llama-cli \
    -hf unsloth/Muse-Glimmer-30B-GGUF:UD-Q4_K_XL \
    --temp 1.0 --top-p 0.95 --top-k 64

Context runs to 131,072 tokens by default.

The arithmetic

Take the monthly API spend on the jobs you would move: drafting, summarising, classifying. The model itself is free. Against that, put the machine if you have to buy one, spread over its life, plus the hours someone spends keeping it alive. A small monthly API line loses that comparison. Steady daily volume is what makes local win, not the licence.

Limitations

• An office Mac runs the small model, not the frontier one. Do not quote Spark Open capabilities to your team and then install Glimmer.
• Quantised output degrades. Hugging Face noted the 2-bit build needs more steering.
• No vendor uptime, no support line, no audit trail unless you build one.
• I have not run Glimmer on a 32 GB Mac myself, so treat every minute figure above as an estimate.
• Local does not mean compliant. Data staying on your disk is one control out of many.

My take

Put it on one machine, give it one boring recurring job, keep the log. If the log stays clean for a fortnight, move that job across and leave everything customer facing on the paid API. Other write-ups sit on the blog.