Cook and Zuckerberg bury the hatchet? Apple Inc. (AAPL.US) and Meta (META.US) team up in a rare move to bring an 816-billion-parameter open-source model "into" the Mac, shifting the landscape of on-device AI

Meta (META.US) and Apple Inc. (AAPL.US) announced a new collaboration on Thursday that will allow developers to conveniently run Meta's new open-weight AI model on Apple Inc.'s high-end desktop computers. According to Meta, at the heart of the collaboration is Muse Spark Open, the company's most powerful open-weight model released under a license permitting limited commercial use. It is a model with 816 billion parameters, employing a so-called mixture-of-experts architecture. As an open-weight model, developers can download the technology's parameters (i.e., weights) from the internet and run the model on a variety of hardware. The two companies said users can network two to four Mac Studio machines together to use Muse Spark for programming, analyzing video, or performing other agentic tasks. The model runs on local devices without needing the cloud, which both enhances the level of data security and saves users on cloud service costs. The collaboration, announced Thursday after months of refinement, signals that the AI boom is reshaping Silicon Valley's longstanding camp alignments. For a long time, Apple Inc. and Meta had been wary of each other, with virtually no public collaboration. In 2021, Apple Inc. introduced privacy changes that significantly weakened digital advertisers' targeting capabilities, pushing relations between the two to a low point. Meta estimated in early 2022 that the change would reduce its revenue by $10 billion that year. Representatives for both Apple Inc. and Meta declined to say whether the Muse Spark Open collaboration signals a deeper strategic partnership between the two. Zuckerberg posted a screenshot on X showing Meta's AI agent running on a Mac. He wrote: "Muse Spark Open officially available for open-weight download today. Thanks to @johnternus and the Apple Inc. team for their collaboration; the model performs excellently on the new Mac Studio." Ternus responded: "Congratulations to the team! It's truly amazing to run a large model of Muse Spark Open's scale and specifications so efficiently on the desktop. Mac Studio continues to push the performance boundaries of on-device AI!" Meta AI chief Alexander's, Inc. Wang said in an interview that many powerful open-weight models are too large to run on most personal computers, but Apple Inc.'s Mac Studios provide users with enough memory to make local deployment of Muse Spark Open practically feasible. The collaboration is both an important show of confidence as Meta restarts its AI push and a potential competitive threat to OpenAI and Anthropic. At the same time, it sends a signal that Apple Inc.'s AI chip capabilities are steadily strengthening. Currently, NVIDIA Corporation (NVDA.US) has long dominated the AI chip market thanks to its widespread deployment in large data centers. "We are thrilled to work with Apple Inc. to bring top-tier agentic capabilities to users for the first time on Mac Studio desktop devices powered by the M5 Ultra chip," Wang said, noting that developers can download the model from platforms such as Hugging Face and run it on Apple Inc.'s custom chips without additional adaptation. Muse Spark Open has been optimized and accelerated for Apple Inc.'s hardware. Apple Inc. simultaneously released new benchmark data for the chip and published technical documentation detailing how the MLX AI framework splits open-source models and runs them in a distributed manner across multiple machines. However, the Mac Studio is not cheap, and the AI model has hard requirements for Apple Inc.'s top-tier chips and large memory capacity. Currently, a Mac Studio equipped with the required M5 Ultra chip and 256GB of memory starts at at least $10,799. A higher-spec version with 512GB of memory will go on sale later this month. In addition, users need to purchase multiple machines to network together. Consensus in the open-source AI field The two companies are collaborating in the open-source AI field, which has become a point of controversy in Washington because politicians worry that China's free open-source models could undermine American AI. Although Meta's AI strategy in its early years centered on the open-weight Llama series, under the leadership of Alexander's, Inc. Wang the company has since shifted its focus to proprietary closed-source models. Alexander's, Inc. Wang joined Meta last year, when the company invested more than $14 billion in his former startup Scale AI. Although Apple Inc. and Meta (then named Facebook) collaborated more than a decade ago on iPhone feature expansion, the two have not reached any public collaboration since Apple Inc. removed system-level Facebook integration from the iPhone in 2017. "Most of the conflict between the two sides existed before I joined. The entanglement between Meta and Apple Inc. goes back a long way," Wang said, noting that one member of his team AI researcher and former Apple Inc. director Daniel Gross was one of the reasons the collaboration came together. Gross, like former GitHub CEO Nat Friedman, was a key recruit by Zuckerberg last year when he overhauled the company's AI strategy. There have also been sharp verbal exchanges between the leaders of the two sides. Cook once took a veiled jab at Zuckerberg during the 2018 Cambridge Analytica scandal, saying when he was still Apple Inc. CEO that "I wouldn't get into that situation." Cook frequently criticized Meta for harming user privacy. (Ternus succeeded Cook as Apple Inc. CEO last month.) Zuckerberg has also publicly attacked Apple Inc., saying on a podcast last year that Apple Inc. "hasn't invented anything great since the iPhone." Zuckerberg has long wanted to build a platform fully controlled by Meta, freeing it from dependence on Apple Inc. Ternus, a hardware engineer by background, has already begun adjusting Apple Inc., including changing its product release cadence. And Meta's internal developers are likewise fond of Apple Inc. devices. Earlier this year, Meta purchased hundreds of Macs for its superintelligence team to conduct local AI and agent-related experiments. Apple Inc.'s multiple AI partners Meta is only the latest AI model developer to partner with Apple Inc. In 2024, Apple Inc. announced a partnership with OpenAI to integrate ChatGPT into Siri, though OpenAI said the integration "performed far below expectations" and did not bring OpenAI user growth. That claim came from a document OpenAI submitted in its legal dispute with Elon Musk's xAI. Apple Inc. sued OpenAI in July, alleging theft of trade secrets, and OpenAI denied the allegation. Earlier this year, Apple Inc. announced an agreement with Alphabet Inc. Class C (GOOGL.US) to integrate the Gemini model into Apple Intelligence, while also adopting Alphabet Inc. Class C cloud infrastructure. Nevertheless, Apple Inc.'s AI strategy is still criticized by investors and customers, partly because the new version of Siri it released last month is less capable than the chatbots and agents of leading labs. Meanwhile, Meta is going all out to compete among top AI labs, but has struggled to keep pace with rivals, and its business still depends almost entirely on advertising. Some recent developments, such as Meta's Muse personal AI agent app and its latest closed-source foundation model Muse Wonder released earlier this week, have begun to excite investors optimistic about new business opportunities for Meta AI. Open source remains an important part of Zuckerberg's AI strategy. In August, Meta launched a small open-weight AI model called Muse Glimmer, and Zuckerberg promised to eventually release larger and more powerful products. Muse Spark Open's performance falls short of the closed-source product Muse Wonder, but Wang said it is enough to attract more programmers to upgrade their hardware and give it a try. "These are not the strongest models we have developed or released," Wang said. "But we believe many companies and developers will use and modify this model on a small scale, and then expand deployment later."