Apple and Meta Team Up on Muse Spark Open: What the Mac Studio Partnership Means

Apple and Meta have announced that they worked together so Meta's new open-weight model, Muse Spark Open, runs locally on clustered Mac Studio computers. The news, first reported by CNBC on October 15, 2026, pairs two companies that spent years avoiding public partnerships. Meta CEO Mark Zuckerberg posted on X that the model "runs great on new Mac Studios thanks to work with @johnternus and Apple," according to CNBC, and Apple CEO John Ternus replied with congratulations.

The model news itself is covered in our Muse Spark Open guide and the benchmark detail in our write-up of Apple's MLX and RDMA research. This post is about the third thing: the alliance, what it costs, and which claims are still loose.

TL;DR: what was announced

What exactly are the two companies doing together?

CNBC describes the collaboration as the result of months of work. The visible pieces are three. First, Meta tuned Muse Spark Open to run faster on Apple's hardware. Second, a version of the model is downloadable from sites like Hugging Face without extra configuration for Apple silicon, per Wang's comments to CNBC. Third, Apple published new benchmarks for its chips and technical documentation about how its MLX framework splits open models across several machines.

That third piece is the one developers can verify today. Apple Machine Learning Research reports about 36.7 tokens per second decoding Muse Spark Open on four M5 Ultra Mac Studios linked by RDMA over Thunderbolt 5. Those are Apple's numbers on Apple's hardware, and our earlier post covers the tables and the caveats, including batch size 1 and 4-bit weights.

The practical promise, in Wang's words relayed by CNBC, is to "bring leading Agent capabilities to your desktop for the first time" on M5 Ultra Mac Studios. He also argued that many strong open-weight models are too big for most personal computers, but Mac Studio memory makes this one practical.

Why this pairing is surprising

The history matters because it explains why the story is being covered as more than a driver update. CNBC recalls that Apple's 2021 privacy changes limited ad targeting and that Meta estimated in early 2022 the change would cut its revenue that year by $10 billion. It also notes Tim Cook's 2018 jab at Zuckerberg during the Cambridge Analytica scandal and Zuckerberg telling Joe Rogan last year that Apple "hasn't invented anything great since the iPhone." The two companies have not struck a public partnership since 2017, when Apple removed system-level Facebook integration from the iPhone.

What changed? CNBC points to leadership: Ternus succeeded Cook as Apple CEO last month (see our coverage of Cook stepping down and Ternus taking over). Wang also credited AI researcher Daniel Gross, a former Apple director now on his team, as one reason the collaboration happened. And Meta's own people like the hardware: CNBC reports the company bought hundreds of Mac computers so its Superintelligence team could experiment with local AI and agents earlier this year.

What it costs: the part the announcement leaves out

A home workstation running a local model, illustrating the hardware cost of an Apple Meta Muse Spark Open setup

Local does not mean cheap. CNBC's figures: a Mac Studio with the M5 Ultra chip and 256GB of memory costs at least $10,799, a 512GB version is due later this month, and customers need to pair two to four machines. Apple's research post, as we noted before, does not publish prices, so this is CNBC's reporting.

Doing the arithmetic from CNBC's number alone, two of the cheapest qualifying machines would be at least $21,598 before cables and tax, and four would be at least $43,196. That is our calculation from one reported starting price, not a quote. The 512GB configuration will cost more, and we have not seen its price.

Compare that with the alternative the companies are pitching against: paying for tokens in the cloud. CNBC says the companies argue local operation adds a layer of security and saves users money. Whether the savings are real depends on volume. A team running a few million tokens a day will probably never recoup the hardware; a team with sensitive code that cannot leave the building has a different calculation.

The memory numbers do not agree

Here is where careful readers should slow down. TechCrunch reports that running the model locally requires "at least 1 terabyte of RAM, per Engadget." CNBC says users need two to four Mac Studios, with 256GB or 512GB each. Apple's own research post, which our earlier article summarized, puts the 4-bit NVFP4 weights at about 493GB, which is why a single 512GB machine is tight and a cluster is used.

These can all be true at once, because the requirement depends on the checkpoint:

• BF16 weights: the Hugging Face launch guide cited in our Muse Spark Open post gives roughly 1.5 TB, far beyond any single Mac.
• 4-bit NVFP4: about 493GB in Apple's account, split across two to four machines.
• 2-bit quantization: the same Hugging Face guide reports a 2-bit llama.cpp quant using about 304GB on a 512GB Mac M3 Ultra at 18 tokens per second.
• KV cache: a full million-token context adds a very large cache on top of the weights, per the Hugging Face arithmetic we summarized.

So "at least 1 TB" is a reasonable floor for higher-precision or full-context use, while a 4-bit cluster of two to four machines fits Apple's setup. We could not find a single official statement that reconciles every figure, so treat any single number as conditional on quantization and context length.

What is claimed and what is verified

What Meta says the model can do

TechCrunch relays two demos from Meta. The company paired the model with Muse Code, its AI coding terminal, and ran it on agentic workflows including coding fixes for a Mastodon client ("bug fixes, new features, [and] full build"), and used it to run DaVinci Resolve, reviewing and cropping footage. TechCrunch also says the model is available through a number of API platforms now. These are Meta's own demonstrations, not independent tests.

CNBC adds that Wang described the model as not Meta's most capable: "These aren't the most capable models that we've developed or released," he said, positioning it to attract developers who tinker at small scale and then scale up. Meta's proprietary Muse Wonder, covered in our Muse Wonder launch post, is the more capable closed model.

How the deal fits Apple's wider AI strategy

CNBC frames Meta as "just the latest AI model developer to team up with Apple." Its list: the 2024 ChatGPT-in-Siri arrangement, which OpenAI said in a filing "dramatically underperformed"; a deal earlier this year to use Google's Gemini models in Apple Intelligence and Google Cloud infrastructure; and Apple's July lawsuit against OpenAI alleging trade secret theft, which OpenAI denies. Our posts on Apple's Siri delegation to outside models and the OpenAI agreement over Apple trade secrets cover that thread.

The pattern: Apple sells the box, and increasingly lets someone else supply the weights. For a hardware company with a Siri that CNBC says is less advanced than leading chatbots, being the best place to run a strong open model is a coherent play. It also lines up with Apple's earlier push on Mac Studio M5 Max and Ultra for local AI.

There is a friction point too. In early October we covered a dispute in which Apple moved to tighten Full Disk Access for AI agents after Meta's Muse Messages behavior, and Meta's agent app is already on the Mac through Muse for Mac. A friendlier tone on local models does not erase the platform-control questions underneath.

What this means for what you build or pay

On-device AI concept showing a model running locally rather than in the cloud for Mac Studio developers

If you are a developer choosing where to run agents:

1. Privacy-bound teams (legal, health, finance code) get a credible all-local option with a frontier-sized open model, at a five-figure hardware price.
2. Everyone else should keep doing the math on hosted APIs. At batch size 1 and roughly 37 tokens per second, a cluster serves one person well; it is not a multi-user inference server.
3. Smaller budgets can look at the smaller open model, Muse Glimmer, which at 30B is a far lighter lift, or at the 2-bit quant route that fits a single 512GB machine at lower quality.
4. Check the license before shipping a product. CNBC says "limited commercial use"; AWS calls it the Muse Model License with a usage policy, per our earlier post.

How to verify this yourself

• Read Apple's MLX distributed inference post for the benchmark tables, versions (MLX v0.30.1 or later, macOS 26.2 or later) and the mlx.launch setup.
• Read the CNBC report for the direct quotes and the price.
• Check Apple's online store for current M5 Ultra Mac Studio memory options and prices, since CNBC's figure is a "at least."
• Compare the weights you download (BF16, NVFP4, GGUF quant) against the memory you have, including KV cache for long contexts.

Bottom line

The substance is a modest, concrete collaboration: tuned weights, a documented way to split a huge model across Macs, and published benchmarks. The symbolism is bigger: two rivals publicly standing next to each other because local, open-weight AI gives each something. What remains unproven is independent performance, real-world memory needs across configurations, and whether either side wants this to go further. Both declined to say.

Figures and quotes are as reported on October 15, 2026 by CNBC, TechCrunch and Apple Machine Learning Research, and may change.

Related reading

• Muse Spark Open: specs, licence and how to run it
• Apple shows 800B-parameter models on four Mac Studios with MLX and RDMA
• Meta Muse Wonder full launch and Muse Code
• Meta Muse Glimmer: a 30B open-weight agentic model
• Tim Cook steps down, John Ternus becomes Apple CEO
• Apple Mac Studio M5 Max and Ultra for local AI