Meta and Apple Team Up: Muse Spark Open Now Runs Locally on Mac Studios (Here’s the Catch)

Meta and Apple, two companies that have spent years sniping at each other, just shipped something together. On Thursday, October 15, Meta released Muse Spark Open, a huge open-weight AI model, and Apple published the tooling and benchmarks to run it on your own desk using Mac Studio computers. CNBC calls it “an unlikely alliance between two longtime rivals.”

Here’s what’s confirmed, what it costs, and whether it matters to you.

What actually got announced

Meta says Muse Spark Open is “the most capable US open-weight model to date.” It’s an 816-billion-parameter mixture-of-experts model with 42 billion parameters active at a time, per Meta’s launch post. “Open-weight” means you can download the model files and run them yourself, instead of only renting access through someone’s cloud. CNBC notes the license allows limited commercial use, so read it before you build a business on it.

Meta’s post says the model “was optimized to run on Mac Studio with M5 Ultra” using Apple’s MLX framework, and calls it the first model in its class to run on a desktop using two of those machines.

Apple backed that up with its own machine learning research post. Apple says that on a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second, which it describes as “well above average human reading speed.” Apple credits a year of MLX work: support for the Neural Accelerators in M5 chips, fast links between Macs using RDMA over Thunderbolt, and better handling of mixture-of-experts models.

Both bosses weighed in publicly. According to CNBC, Mark Zuckerberg posted that the model “runs great on new Mac Studios thanks to work with @johnternus and Apple,” and Apple CEO John Ternus replied that “Mac Studio keeps pushing the boundaries of on-device AI.”

The catch: “local” doesn’t mean “on your laptop”

This is the part the headlines skip. The model is too big for one Mac. Meta’s hardware docs list the Apple requirement as two Mac Studios with M5 Ultra and 512GB of unified memory each, or four with 256GB each.

Meta’s setup guide adds the rest:

• macOS 26.2 or later on every Mac
• MLX 0.30.1 or later
• A direct Thunderbolt 5 cable between every pair of Macs (four Macs means each one connects to the other three)
• Room on every Mac for the full model files
• About four hours to get it all running

Once it’s up, you get an OpenAI-compatible API on your own network, so tools that already talk to OpenAI-style endpoints can point at your Macs instead.

And the price? CNBC reports a Mac Studio with the required M5 Ultra and 256GB of memory starts at $10,799, and the 512GB version ships later this month. You need at least two. So this is a five-figure setup before you’ve typed a prompt.

Why it matters for everyday users

Honestly, most people won’t run this model at home. But the direction matters:

• Privacy: A model running on your own hardware doesn’t send your prompts or files to anyone’s cloud. The companies pitched local use as “adding a layer of security,” per CNBC.
• Trickle-down: Apple’s MLX improvements aren’t Muse-only. Apple’s own post tested smaller models too, Mistral Medium 3.5 (128B) and Qwen 3.8 (27B), starting from a single Mac Studio. The same software work helps smaller models run better on smaller Macs.
• Competition: CNBC says the move could challenge OpenAI and Anthropic, and signals Apple’s growing strength in AI chips, a market Nvidia dominates. More choice usually means better prices.

Why it matters for small businesses

If you’re a small shop, here’s the honest math. Two Mac Studios are a big purchase, but they’re a one-time purchase. If you’re paying a lot every month for cloud AI on sensitive work like client contracts, health records, or source code, a local setup could make sense, and the data never leaves your office.

If you aren’t, there’s a cheaper path. Meta’s launch post says Muse Spark Open is also available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Try it there first. If it’s good enough for your work, then think about hardware.

Two practical tips:

• Check the license. CNBC describes it as allowing limited commercial use. Know your limits before you ship a product on it.
• Know what you’re getting. Meta AI chief Alexandr Wang told CNBC, “These aren’t the most capable models that we’ve developed or released.” Meta’s top proprietary model, Muse Wonder, is separate and isn’t part of this.

Is this a bigger Apple-Meta deal?

Not that anyone’s confirmed. CNBC reports that representatives from both companies “declined to say if the Muse Spark Open collaboration was the start of a deeper relationship.” There’s no announced plan to put Muse in Siri, Apple Intelligence, or the iPhone. Apple already works with Google’s Gemini for Apple Intelligence, CNBC notes. Treat any talk of a wider deal as speculation until one of them says otherwise.

Still, it’s notable. CNBC says the two haven’t had a public partnership since 2017, and the thaw comes right after Ternus took over from Tim Cook.

The bottom line

Meta and Apple really did team up, and you really can run a near-trillion-parameter model on Macs at your desk. You’ll just need at least two very expensive ones. For most people, the takeaway is simpler: local AI keeps getting more capable, and the software making that possible is getting better for everyone.

Sources: Meta AI Research, Apple Machine Learning Research, Meta Model API docs, CNBC.