Every sitcom has the mismatched roommates who bicker constantly and somehow split the rent anyway. The Odd Couple ran five seasons on that premise. On October 15, Apple and Meta announced their own version: Meta’s new Muse Spark Open AI model now runs on Apple’s M5 Ultra Mac Studios, thanks to work between the two companies. For context, these are two companies with a long history of friction over privacy, advertising, and who controls the platform, and per MacDailyNews, they had not publicly partnered since around 2017.

First, what is this thing. “Open-weight” means Meta publishes the finished model so developers can download it and run it themselves instead of renting it from someone’s cloud. (AppleInsider notes Meta’s own license still has commercial limits, so “open” isn’t the same as “free for all.”) Muse Spark Open has 816 billion parameters, which you can think of as the dials the model tuned during training, though only 42 billion are active at a time for each chunk of text it writes, per PPC Land. That’s a gigantic office building where only the relevant departments pick up the phone.

Here is where the roommate thing gets literal. The model is too big for one computer. AppleInsider reports the compressed model alone takes about 493GB before it starts doing any work, and Meta’s own setup uses two Mac Studios with 512GB of memory each, linked so they behave, in Meta’s words, “like one system with a terabyte of memory.” Apple updated its open-source MLX software so the machines can split the load. The AI doesn’t live on your desk. It lives on your desk and your other desk, and they have to get along.

Then comes the rent. MacDailyNews reports a Mac Studio with the M5 Ultra chip and at least 256GB of memory starts at $10,799, and PPC Land notes that the 512GB version Meta used hadn’t been priced yet. Two of the cheaper ones already come to about $21,600. Running AI locally is supposed to offer advantages in security, cost, and less reliance on cloud data centers, according to that same report, and the cost part is a bold claim to make about a setup that costs more than a decent used car. Local AI used to mean “runs on my laptop.” Now it means “runs in a room, and the room has a lease.”

How good is it? Meta’s demos had the model fix bugs, add features, and build an update to an open-source iPhone client for Mastodon, and review raw video in DaVinci Resolve to cut vertical clips. AppleInsider cautions the demos don’t show how reliably it handles other work, and PPC Land notes Meta’s benchmark scores haven’t been independently reproduced. A Meta executive also said the model is “not Meta’s most capable,” which means the two-computer, five-figure setup is running the junior varsity.

If you want to try it without remortgaging, the model is on Hugging Face, and AppleInsider says developers without the hardware can use hosted API providers. Everyone else can admire it from the cloud, where it was apparently going to live anyway.

Representatives from both companies declined to say whether this signals a broader relationship, which is exactly what every roommate says before signing the second lease.

Sources: AppleInsider — Unlikely Apple & Meta Team-Up Brings Powerful & Local AI to Mac Studio (https://appleinsider.com/articles/26/10/15/unlikely-apple-meta-team-up-brings-powerful-local-ai-to-mac-studio); PPC Land — Apple and Meta Team Up to Run Meta’s 816B Muse Model on Two Mac Studios (https://ppc.land/apple-and-meta-team-up-to-run-metas-816b-muse-model-on-two-mac-studios/); MacDailyNews — Apple and Meta Partner to Run Massive Muse AI Model Locally on Mac Studios (https://macdailynews.com/2026/10/15/apple-and-meta-partner-to-run-massive-muse-ai-model-locally-on-mac-studios/)