Meta Just Gave Away an 800-Billion-Parameter AI Model for Free. To Run It You Need a Terabyte of RAM and About $38,000 of Macs.

YOU get the weights! YOU get the weights! EVERYBODY gets the weights!

Meta — working with Apple, which is a sentence I did not expect to type — just put an 800-billion-parameter model called Muse Spark Open on Hugging Face. Not an API. Not a waitlist. The actual weights, tuned to run locally on Apple silicon, free to download. Meta is calling it the most capable US open-weight model to date, and says it beats the Chinese open models GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning.

That is a genuinely big deal. For years the best models lived behind somebody's billing page. This one lives on your hard drive.

So look under your seat. Go on.

Oh. Oh honey.

It needs one terabyte of RAM.

Not a typo. To run this at home you need two Mac Studios — 512 gigabytes each — wired together as a single machine. Or four M5 Ultra Mac Studios at 256 gigabytes apiece, which runs you somewhere around thirty-eight thousand dollars.

It works because Apple's unified memory architecture lets the M5 Ultra's integrated GPU reach up to 512GB per Mac Studio. Beautiful engineering. Also: thirty-eight thousand dollars.

So it's free the way a puppy is free. The puppy costs nothing. The puppy also needs a yard.

I want to be fair here, because I love a giveaway and this is one. Open weights mean researchers can inspect it, universities can study it, and nobody can quietly swap the model out from under you or change the price next quarter. Somebody with a lab and a budget is going to do remarkable things with this. Those are real, permanent wins, and they don't disappear because the hardware bill is steep.

But let's tell the truth about who walked out with the car keys. "Open" and "accessible" turned out to be two different words. The model is open to everyone. It is runnable by people who can drop the price of a nice car on desktop computers.

What this means for the rest of us

Here's where I'd land it.

If you can't host an 800-billion-parameter model in your spare room — and I'm going to guess that's most of us — the practical version of this freedom isn't owning one model. It's not being captive to one. The leaderboard rotates every few weeks. Today it's Meta. Last week somebody else. The person who bet everything on a single lab's model is always the last to find out it slipped.

So for anything that actually matters, don't take one model's word for it. Ask several of the strongest ones the same question and watch what they do. Gangsta AI puts your question to the most powerful models at once with an adjudicator weighing them against each other — so when they agree you've got something solid, and when they split you've found the exact spot where one of them is wrong. That's a kind of access that doesn't require a terabyte of anything.

Meta gave away something genuinely valuable yesterday. Just read the system requirements before you clear off the desk — and get a second opinion before you bet on any one of them.