How to run AI locally on a Mac with Meta’s Muse Spark Open

You can now run AI locally on a Mac with a model close to the big cloud ones, but not on a laptop. Meta’s new Muse Spark Open runs on two M5 Ultra Mac Studios with 512GB of memory each, and your prompts and client files never leave the office.

Meta released it on October 15. It’s free to download and built for coding and agent work.

Meta also tested it on Macs on purpose, which is new for a model this size. Here’s Meta’s announcement.

What did Meta release?

• The model. 816 billion parameters, with 42 billion active at a time. It reads up to 1 million tokens at once, so it can take in a big pile of documents in one go.
• What it’s for. Coding and long, multi-step agent tasks. Meta showed it making changes to an iOS app and driving DaVinci Resolve.
• Where to get it. The files are on Hugging Face. OpenRouter, Together AI, Fireworks, Groq and others also host it if you just want an API.
• The claim. Meta calls it the most capable US open-weight model to date. That comes from Meta’s own testing. Nobody has checked it independently yet.

“Open-weight” means you download the actual model and run it yourself. Claude and ChatGPT never give you that. You rent them over the internet.

What Mac do you need to run it?

You need about a terabyte of memory.

Meta’s Mac cookbook lists two setups:

1. Two Mac Studios with the M5 Ultra chip and 512GB of memory each.
2. Four Mac Studios with 256GB each.

Every Mac needs macOS 26.2 or later. Each pair connects with a Thunderbolt 5 cable.

Engadget puts the four-Mac setup at roughly $38,000. Apple hasn’t priced the 512GB Mac Studio yet and says it ships in late October.

So your MacBook won’t run it. Every Mac setup Meta lists uses two or four Mac Studios.

Why run AI on your own Mac at all?

Privacy. When you paste a client contract into ChatGPT, it goes to OpenAI’s servers. Run the model locally and it stays on hardware you own.

That matters most if you handle legal, medical, financial or personal data. Same if your client agreements limit where their data can go.

The other reason is cost at volume. You pay for the Macs once, then nothing per message. For most small teams, a subscription still comes out cheaper.

How does it compare to Claude and ChatGPT?

• Setup. Claude and ChatGPT: open a browser tab. Muse Spark Open on your Macs: link the machines, then run a few terminal commands.
• Where your data goes. Their servers, or your office.
• Cost. A monthly bill, or tens of thousands up front.
• Quality. No independent comparison with Claude or ChatGPT yet. Don’t switch on Meta’s numbers alone.

My take: keep Claude for daily work. I still use it for things like building dashboards from a spreadsheet. Look at local AI when data rules stop you from using cloud AI at all.

How do you try it?

There are two routes.

No new hardware: use the API. It’s live on OpenRouter and Together AI. That’s the fastest way to see if it’s any good at your work.

But your data goes to that provider, so you lose the privacy win. Test it with nothing sensitive.

On your own Macs. Once the Macs are linked and RDMA is switched on, Meta’s cookbook goes like this:

1. Install MLX 0.30.1 or later and mlx-lm on every Mac:

pip install -U mlx-lm

2. Set up passwordless SSH from the main Mac to the others.
3. Create a hostfile.json that lists every Mac. The mlx.distributed_config tool can generate it.
4. Start the server from the main Mac:

mlx.launch --hostfile hostfile.json --backend jaccl --env MLX_METAL_FAST_SYNCH=1 -- \
  mlx_lm.server --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --port 8080

That gives you an API on port 8080 in the same format as OpenAI’s. Tools that already talk to ChatGPT’s API can point at your Macs instead.

What it can’t do

It won’t run on the Mac you have now, unless that’s a maxed-out pair of Studios.

The license also has strings. Any product built on it has to show “Built with Muse Spark Open.” Selling access as an API? Past $20 million a year in revenue, you need a separate deal with Meta.

Using it inside your own business has no revenue cap. Read the full license before you build on it.

What to do now

If client data has kept you away from AI, write down the tasks you’d hand to a private model. This AI task audit is a good way to sort them.

If the list is long, local AI is worth pricing out. If it’s short, the API route is enough to test.

FAQ

Can I run AI locally on a MacBook?

Not this model. Muse Spark Open needs about a terabyte of memory, which means two or four Mac Studios. Smaller open models do run on MacBooks.

Is Muse Spark Open free for business use?

It’s free to download and use under the Muse Model License. Products built on it must show “Built with Muse Spark Open.” If you sell it as an API and pass $20 million a year, you need a separate deal with Meta.

Is running AI locally more private than ChatGPT or Claude?

Yes, if it runs on your own hardware. Your prompts and files stay on your machines. Using it through a hosted API like OpenRouter doesn’t give you that.

Is Muse Spark Open better than Claude?

Nobody knows yet. Meta says it beats other open-weight models, but there are no independent tests against Claude so far.

What Mac do I need for Muse Spark Open?

Two M5 Ultra Mac Studios with 512GB each, or four with 256GB. All on macOS 26.2 or later, linked with Thunderbolt 5.

Feel free to book a call if you need help.