Apple and Meta, whose relationship has been shaped since 2021 by Apple's tracking restrictions and Meta's response to them, have worked together for several months to make Meta's largest open AI model run on Apple's desktop computers. Mark Zuckerberg made the collaboration public on October 15, 2026, writing on X that Muse Spark Open "runs great on new Mac Studios thanks to work with @johnternus and Apple." Apple chief executive John Ternus replied the same day with congratulations.

In Short

Apple and Meta, two companies that have spent years fighting over privacy and ad tracking, worked together so Meta's big new AI model can run on two of Apple's Mac Studio computers. That matters because the model is from the same family that powers Meta's Muse agent, which now reaches into advertisers' ad accounts, and because running it on a desk means Meta never sees what people type into it. What changes for now is limited: there is no money involved, no Siri or iPhone deal, and the Mac Studio model needed for the setup was not yet on sale.

What the two companies said

The public record from the companies themselves is two posts. Zuckerberg's, captured with a timestamp of 7:01 PM on October 15, 2026, reads in full: "Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple." By the time of the capture it had 4.4 million views, 1,100 replies, 2,500 reposts and 31,000 likes.

Ternus answered within the same thread. "Congrats to the team! 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!" he wrote. The reply drew 1.2 million views and 14,000 likes. It is the only statement from Apple in the material reviewed, and it came from the chief executive himself, six weeks into the job: Ternus succeeded Tim Cook on September 1, 2026, a handover PPC Land reported when Apple disclosed it alongside a record $111bn quarter in May.

More detail came from Meta's side. According to 9to5Mac and MacRumors, Meta's AI chief Alexandr Wang told Axios that the two companies had been working together for several months, and credited Daniel Gross, a former Apple executive now at Meta, with a large role in bringing them together. Wang said, according to Axios, that the companies would keep competing in some areas but partner where it made sense. He also said Meta has no control over or visibility into how developers use the model once they download it.

No financial terms were disclosed, and nothing in the reporting suggests any exist. Apple did not issue a newsroom post. According to 9to5Mac, Apple's Machine Learning Research site published an article on distributed inference with MLX, its framework for Apple silicon, describing local experimentation with the latest models in a private environment.

What each side brought

The division of labour is visible in Meta's research blog. Meta supplied Muse Spark Open, a natively multimodal model with 816 billion total parameters, 42 billion of them active per token, and a context window of 1 million tokens, aimed at coding and agentic work. Apple supplied the hardware path: the M5 Ultra chip, whose integrated GPU can address up to 512GB of unified memory per Mac Studio at 1.2TB per second of bandwidth, and MLX, which lets two linked machines behave, in Meta's description, like one system with a terabyte of memory.

That terabyte matters because of how the model is built. Every one of the 816 billion parameters has to sit in memory even though only about 5.1% of them do work on a given token, by PPC Land's calculation. At 16-bit precision the weights alone would take roughly 1.6TB, also PPC Land's calculation, so the two-machine configuration implies reduced-precision weights. Meta's blog does not say which precision its demonstration used.

Meta showed the setup running Muse Code, its coding tool, on-device. According to the blog, it fixed bugs, added features and completed full builds of Ice Cubes, an open-source Mastodon client for iOS, and reviewed and cropped footage in DaVinci Resolve. Both are workloads Apple sells the Mac Studio for.

The Mac is not an exclusive home. Meta also published the weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and lists hosted access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, with optimisations for AMD, Intel and NVIDIA hardware alongside Apple's. Apple's role is to be the desktop option among several data-centre routes.

Why each company wants it

The arrangement solves a problem both companies have. Because the model runs on the user's own machine, Meta receives no prompts and no usage data, which, as Axios observed, sidesteps the privacy disputes that have divided the two. For Apple, that is the only kind of Meta AI product it can promote without contradicting its own positioning on privacy. For Meta, it is a showcase for its first downloadable model since it moved its Muse line behind an API.

That move came on July 9, 2026, when Meta opened Muse Spark 1.1 to developers through the Meta Model API rather than as a file anyone could download. Meta framed the new open release as a US answer to Chinese open-weight models; according to Engadget, the company said the US should not let Chinese labs lead in open-source AI. The pitch has a commercial target. When the European search engine Ecosia replaced Mistral this month, it chose open-weight models from Qwen, GLM and Kimi, all from Chinese developers.

Apple, for its part, gets an expensive model to sell machines. The 512GB Mac Studio was not on sale when the post appeared. According to Engadget, Apple had not announced its price or ship date, beyond an expectation of late October. MacRumors noted the 256GB model already costs more than $10,000, and Engadget, citing Technology.org, estimated a four-machine 256GB alternative at roughly $38,000. The "right from your desk" setup in Ternus's reply is, in practice, a workstation purchase for developers and companies.

Five years of friction

The reason a two-line exchange on X drew attention is the history behind it. According to Axios, after the Cambridge Analytica scandal in 2018 Cook said he would not have been in Zuckerberg's situation, and Zuckerberg called the remark extremely glib. The deeper break came with App Tracking Transparency, which Apple introduced in April 2021 with iOS 14.5. Opt-in rates of 11% to 15% followed, cutting into the signal Meta's ad systems used to target and measure campaigns on iPhones. Axios recalled a Wall Street Journal report that Zuckerberg told employees in 2021 that Facebook needed to inflict pain on Apple.

The conflict moved to Brussels. Under the Digital Markets Act, Meta filed 15 interoperability requests for access to iOS technologies including AirPlay, App Intents, CarPlay, iPhone Mirroring and messaging services, and Apple argued in December 2024 that granting them all could expose data it had chosen not to access itself. According to Benzinga, Apple also turned down a Meta proposal in 2024 to bring its Llama chatbot to the iPhone.

The friction did not stop this year. Apple told Mac developers on October 2 that Full Disk Access would require very explicit user action, warning that the risks would grow as AI agents became more autonomous. Its notice named no company, but it arrived amid a dispute over Meta's Muse agent after a columnist reported that Muse had synced about 187,000 lines of his Messages database; Meta disputed that account. Thirteen days later, Apple's chief executive was congratulating Meta on X.

Apple's other AI partner

The collaboration sits beside a much larger one. Apple's foundation models are built with Google: PPC Land reported the arrangement on January 12, 2026, confirmed at WWDC on June 8 when Craig Federighi described a deep collaboration with Google's Gemini models, and shipped in September when Siri AI went live outside the EU and China.

The two relationships are different in kind. Gemini technology sits inside Apple's own products and runs on Apple's servers through Private Cloud Compute. Meta's model sits outside them, as third-party software that happens to run well on Apple hardware, in the same way that Google's own agent came to macOS in June for Ultra subscribers. Axios identified the larger step as Apple adding Meta's models to Apple Intelligence, and reported no plan to do so. Nothing in the posts or the reporting mentions the iPhone, Siri or the App Store.

What it means for advertisers

For the advertising market, the significance is less the model than the relationship around it. The Muse family is now tied into Meta's ad business: the consumer agent went live in the US on September 8, and three weeks later Meta connected it to ad accounts and 15 business tools. According to Benzinga, Muse competes with Apple directly in shopping, travel bookings and payments.

A working channel between the two companies is new. Since 2021, Apple's decisions that affected Meta's advertising, from tracking prompts to interoperability under EU law, reached Meta as rules to comply with or contest. A collaboration where both sides benefit, and where Meta collects no data, does not change App Tracking Transparency or any iOS rule. It does show Apple's new chief executive endorsing a Meta product in public, six weeks after taking over.

The limits are just as clear. The open model's value to agencies and ad tech vendors is that it runs outside Meta's systems, so client data need not pass through Meta's API. That same property means it produces no advertising signal for Meta. The commercial upside for Meta lies elsewhere: in developer adoption of its model family, and in the reported Muse Model License terms. According to Digital Applied's summary of that licence, any product built on the model must display "Built with Muse Spark Open," and hosting providers with more than $20 million in revenue need a separate agreement with Meta. That summary was not checked against the licence text for this article.

Running a model of this size on a desk also puts large LLM inference within reach of teams that would not send sensitive data to a cloud provider, which is the case Apple's research site made for MLX. Whether that matters for agentic AI tools built for marketing depends on whether those teams buy $20,000 to $38,000 in hardware to do it.

How the model compares

Benchmarks are secondary to the partnership story but frame its stakes. Engadget reported Meta's claim that the model beats Zhipu's GLM 5.3 and Moonshot's Kimi K3 on agentic tasks and visual reasoning. Figures reported by Digital Applied and Implicator from Meta's model card are more mixed: 89.2 on the MCP-Atlas tool-use test, but 76.4 on Terminal-Bench 2.1 against 85.0 for Kimi K3 and 83.9 for GLM 5.3. These are company-run results, not independently reproduced at release.

What remains unknown

Several questions have no answer in the record. Neither company has said whether the collaboration extends beyond this release, to future Muse models, to the iPhone or to Apple Intelligence. Apple has issued no statement other than Ternus's reply. The 512GB Mac Studio's price was unannounced. And whether Meta and Apple can cooperate on hardware while Apple's privacy rules continue to shape Meta's advertising revenue is a question both companies left open, with Wang's description of competing in some areas and partnering in others as the closest thing to an answer.

Timeline

Summary

Who: Apple, under chief executive John Ternus, and Meta, under Mark Zuckerberg, with Meta AI chief Alexandr Wang and former Apple executive Daniel Gross, whom Wang credited with bringing the companies together.

What: The two companies worked together for several months so that Meta's open-weight Muse Spark Open model, with 816 billion parameters, runs on two linked 512GB M5 Ultra Mac Studios using Apple's MLX framework. No financial terms were reported, and no Siri, iPhone or Apple Intelligence integration was announced.

When: Made public on October 15, 2026, by Zuckerberg's post on X and Ternus's reply the same day.

Where: On X and Meta's research blog, with the model downloadable from Hugging Face and major cloud catalogues and runnable on Apple's Mac Studio.

Why: Running locally means Meta sees no user data, which avoids the privacy conflict that has divided the companies since App Tracking Transparency in 2021. Meta gets a showcase for its open model against Chinese rivals; Apple gets a reason to sell high-memory Mac Studios. For advertisers, it is the first public cooperation between the company that sets iPhone privacy rules and the company whose ad business those rules hit hardest, without any change to the rules themselves.