Meta and Apple Drop Muse Spark Open, a 42B Open Model

A surprise team-up between old rivals

Meta and Apple have never seen eye to eye on AI, privacy, or much else. That's what makes Muse Spark Open so unexpected.

Meta and Apple have spent years on opposite sides of the tech industry's biggest fights. Apple's 2021 App Tracking Transparency feature cost Meta an estimated $10 billion in 2022 alone, and their approaches to AI couldn't be more different: Meta has burned billions chasing open models, while Apple has moved slowly and focused on privacy.

So when Meta announced Muse Spark Open, an open-weights model built with Apple's hardware team, it caught more than a few people off guard. According to CNET, Meta's chief AI officer Alexandr Wang told Axios, "We're excited for mommy and daddy to be playing nice again." Zuckerberg was more measured, pointing to new Apple CEO John Ternus and the shared goal of on-device AI.

The collaboration doesn't mean the two companies have buried the hatchet on everything. It does mean they've found a patch of common ground: Meta wants its models to run well on as much hardware as possible, and Apple wants its hardware to be the best place to run them.

What Muse Spark Open actually is

Strip away the corporate drama and you get a serious piece of AI engineering aimed at developers and businesses.

Muse Spark Open is an open-weights model, meaning anyone can download it, inspect how it works, and customize it. It has 42 billion active parameters—a measure of how much of the model is engaged during any given task—and a context window of one million tokens. That's enough to process entire books or large codebases in one go.

Meta says the model is particularly good at agentic and coding tasks, which puts it in direct competition with other open-weights models from the likes of Mistral and Alibaba. It's multimodal, so it can handle more than just text. The weights are available now on HuggingFace under Meta's Muse Model license.

For developers, the appeal is control. You can run the model on your own hardware, keep your data local, and avoid paying per-token API fees. The trade-off is that you need the hardware to do it.

The hardware catch: two Mac Studios required

Running Muse Spark Open locally isn't cheap. Apple's own hardware requirements make that clear.

If you want to run Muse Spark Open on an Apple computer, you'll need two M5 Mac Studio Ultras. At $5,499 each, that's a $10,998 investment before you've written a single line of code. The model is optimized for Apple's hardware, but it's not a lightweight affair.

That price tag puts it out of reach for most individual developers, but it's not aimed at them. For businesses that want to keep sensitive data on-premises, the upfront cost can be offset by avoiding cloud API bills. As we've seen with enterprise AI spending, those bills add up fast.

The Mac Studio has become a quiet favorite for AI work since its August release. Earlier this year, when the OpenClaw agent was all the rage, Mac Minis were the go-to machine for many AI enthusiasts. Apple's M-series chips, with their unified memory architecture, are well-suited to running large models locally.

Availability and pricing

Muse Spark Open is available now on HuggingFace under Meta's Muse Model license, free to download. To run it on Apple hardware, you'll need two M5 Mac Studio Ultras, priced at $5,499 each in the US. UK and EU pricing is likely to differ due to taxes and import duties; Apple's regional stores list the Mac Studio at £5,499 and €6,299 respectively, but those figures can change.

The model itself is open-weights, so anyone can use it commercially under the license terms. There's no subscription or API fee if you run it locally. Just bring your own hardware.

Meta and Apple teaming up on an open-weights AI model is a plot twist few saw coming. Whether it's a one-off or the start of a beautiful friendship, it's a reminder that in AI, even rivals can find reasons to work together.