Name: MLX
Author: Apple

Mark Zuckerberg sold Meta‘s new open-weight model in one line on X: “Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple,” the Meta CEO wrote. Apple CEO John Ternus replied in public, and two companies that usually treat each other as rivals spent launch day praising each other’s work.

The weights are real and anyone can download them. The “desktop” in the pitch is something else. Meta’s own recipe pairs two of Apple’s most expensive Mac Studios, each with the largest memory option Apple makes, and on Apple’s US store that memory option is still marked as coming. The model shipped. The exact machine Meta describes for it cannot be ordered yet.

What Meta actually released

Muse Spark Open is the open-weights version of Muse Spark, the model behind Meta’s hosted Muse products, published by Meta Superintelligence Labs on October 15. It is a mixture-of-experts model with about 816 billion parameters, of which about 42 billion work on any given token, and a context window of up to one million tokens. It reads text, images, audio and video and answers in text. Meta trained it for agentic coding, and calls it the “most capable US open-weight model to date,” in a line quoted by The Verge and repeated in Hugging Face’s technical write-up.

It is also a change of direction. Since April, Meta has moved its flagship models behind closed doors, and August’s Muse Glimmer, a 30-billion-parameter model under Apache 2.0, was the small, one-machine member of the family. Spark Open is the full-size model. Its licence is short but not fully open: every product built on it must display the words “Built with Muse Spark Open,” according to Digital Applied’s reading of the Muse Model License, and inference companies selling it as a service above $20 million in revenue need a separate agreement with Meta.

The two-Mac claim, read closely

Meta’s launch post describes running the 4-bit NVFP4 checkpoint, about 530GB, through Apple’s MLX framework on two M5 Ultra Mac Studios with 512GB of unified memory each, or on four machines with 256GB each, linked over Thunderbolt 5 with RDMA. MLX splits every layer across the machines, so each one holds part of the weights and does part of the math, Technology.org explains. A mixture-of-experts design keeps compute low, but not memory: “All 256 experts stay in memory, so deployment needs the memory of an 816B-parameter model,” Hugging Face writes. The cache needed for a full million-token context adds about 266GB on top, by the same write-up’s arithmetic. That is why the recipe starts at a terabyte, not at the file size.

Apple announced the M5 Ultra Mac Studio on August 25 at a starting price of $5,499, and said in the same release that the 512GB model was “coming in late October.” On October 16 the memory selector on Apple’s US store still read “512GB memory option for M5 Ultra coming late October,” and the options it offered stopped at 256GB. Today, the buyable version of Meta’s setup is the four-machine one, which Technology.org prices at roughly $38,000.

No independent speed figure for that setup appears in the launch coverage. The only independent local numbers come from Hugging Face, and they are on previous-generation hardware: on a 512GB M3 Ultra, a 2-bit build ran at 18 tokens per second using 304GB, and a 3-bit oMLX build at about 21 tokens per second in 344GB. Those are heavily compressed versions of the model, closer to a floor than to the experience Zuckerberg describes.

Why Meta and Apple are suddenly friendly

“Congrats to the team!” Ternus replied to Zuckerberg. “Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!” Meta chief AI officer Alexandr Wang told Axios, as The Verge reported: “We’re excited for mommy and daddy to be playing nice again.” Axios also noted the companies had been working together for several months, and Apple published its own technical post on MLX covering Muse Spark Open and two other open models.

Both sides get something concrete. Apple’s August pitch for the new Mac Studio promised that clustered machines could “load the largest and most demanding frontier-class open-weight models available today.” A frontier-scale American model with an Apple recipe on launch day is the workload that pitch needed. Meta gets a desktop story for a model most teams will reach through rented servers instead: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI serve it from day one, alongside the cloud catalogues of Amazon, Google and Microsoft.

The word “US” is doing work

Meta’s benchmark tables also explain the qualifier. On MCP-Atlas, a tool-use test, Meta reports Muse Spark Open at 89.2 against 82.3 for Kimi K3 and 84.2 for GLM 5.3. On Terminal-Bench 2.1, its own figure of 76.4 trails both Chinese models, at 85.0 and 83.9, as Digital Applied lays out. All of those numbers are Meta’s runs, and Technology.org notes that no outside tester has reproduced them yet.

The real test arrives with Apple’s late-October memory option, the first moment anyone outside Meta and Apple can buy the exact two-box setup and publish what it does. Until then, “runs great” is a review written by the company that made the model and the company that sells the machine.

Tags: Apple, meta, Mark Zuckerberg, Open-Weight Models, Mac Studio, Muse Spark Open