Did Cook Zuckerberg complain about the reversal of the story? Apple (AAPL.US) and Meta (META.US) joined forces rarely to “move” the 816 billion parameter open source model to the Mac, changing the end-side AI battle situation

The Zhitong Finance App learned that Meta (META.US) and Apple (AAPL.US) announced a new collaboration on Thursday. Developers will be able to easily run Meta's new open weighted AI model on Apple's high-end desktops.

According to Meta, the core of this partnership is Muse Spark Open, which is the most powerful open weighting model released by the company under a license that allows limited commercial use. This is a model with 816 billion parameters and uses what is known as a hybrid expert architecture. As an open weighting model, developers can download the technology's parameters (i.e. weights) from the internet and run the model on a variety of hardware.

The companies said users can network two to four Mac Studio hosts to program, analyze videos, or perform other intelligent proxy tasks using Muse Spark. This model runs on a local device without the need to go to the cloud, which not only enhances the level of data security, but also saves users the cost of cloud services.

This collaboration, which was officially announced on Thursday after months of refinement, marks that the AI boom is reshaping the camp pattern that has continued in Silicon Valley for many years. Apple and Meta have been wary of each other for a long time and have hardly collaborated openly. In 2021, Apple introduced a privacy adjustment function, which drastically reduced the targeted delivery capabilities of digital advertisers, and the relationship between the two parties reached a rock bottom. Meta estimated in early 2022 that the change would reduce its revenue by $10 billion that year.

The relevant representatives of Apple and Meta both declined to disclose whether the current Muse Spark Open cooperation indicates that the two sides will embark on a deeper strategic cooperation.

Zuckerberg posted a screenshot on the X platform showing the Meta smart agent running on a Mac. He wrote, “Muse Spark Open — officially open for download today. Thanks @johnternus与苹果团队的协作, that model works great on the new Mac Studio.”

Tenus responded by saying “Congratulations to the team! It's amazing to be able to run such a large model of this size and specification as Muse Spark Open so efficiently on the desktop. Mac Studio continues to break through the performance boundaries of end-side AI!”

Alexander Wang, the head of Meta AI, said in an interview that many powerful open weight models are too large to run on most personal computers, but Apple's Mac Studios provides users with enough memory to make it practical to implement Muse Spark Open locally.

This collaboration is not only an important demonstration of confidence in Meta's AI restart layout, but also poses potential competition for OpenAI and Anthropic. At the same time, it also sends a signal that the strength of Apple's AI chips continues to grow. Currently, Nvidia (NVDA.US) has long dominated the AI chip market with extensive deployment in large data centers.

“Together with Apple, we are bringing top intelligent agent capabilities to users for the first time on a Mac Studio desktop device equipped with the M5 Ultra chip, and we are very excited about this.” Wang explained. Developers can download the model from platforms such as Hugging Face and run it on Apple's own chips without additional adaptations. Muse Spark Open has been optimized and accelerated for Apple hardware. Apple simultaneously released new benchmark data for the chip and released technical documents detailing how the MLX AI framework splits open source models and operates distributed across multiple machines.

However, the Mac Studio is expensive, and this AI model has hard requirements for Apple's top chip and large memory capacity.

Currently, the starting price of a Mac Studio with the required M5 Ultra chip and 256GB of memory is at least $10,799. The premium version with 512 GB of RAM will be available later this month.

In addition to this, users also need to purchase multiple hosts for networking.

Reaching consensus in the field of open source AI

The two companies are collaborating in the field of open source AI, and this field has become the focus of controversy in Washington because politicians worry that China's free open source model may weaken American AI. Although Meta's AI strategy was centered around the LLAMA series of open weight models in the early years, the company's focus has since switched to self-developed closed source models under Alexander Wang's leadership. Alexander Wang joined Meta last year when the company invested more than $14 billion in its former startup Scale AI.

Although Apple and Meta (then known as Facebook) collaborated on iPhone extensions more than 10 years ago, the two sides have not reached any public cooperation since Apple removed system-level Facebook integration from iPhones in 2017.

“Most of the conflicts between the two parties existed before I joined the company. “The dispute between Meta and Apple has been around for a long time,” Wang said, pointing out that a member of his team — Daniel Gross, an AI researcher and former Apple director — is one of the reasons for this collaboration. Gross, like former GitHub CEO Nat Friedman, was an important signer for Zuckerberg last year when comprehensively overhauling the company's AI strategy.

There have also been fierce verbal clashes between the leaders of the two sides.

Cook sneaked out at Zuckerberg during the 2018 Cambridge Analysis scandal. At the time, he was still the CEO of Apple, saying “I won't fall into that situation.” Cook has often criticized Meta for compromising user privacy. (Tenus succeeded Cook as CEO of Apple last month.)

Zuckerberg also publicly criticized Apple. As a guest on a podcast last year, he bluntly stated that Apple “hasn't invented anything great since the iPhone.” Zuckerberg has long wanted to build a platform completely controlled by Meta to get rid of its dependence on Apple.

Tenus, a hardware engineer, has begun making adjustments to Apple, including adjusting the pace of product releases. And Meta's in-house developers are also fond of Apple devices. Earlier this year, Meta purchased hundreds of Macs for the super-smart team to conduct local AI and smart agent-related experiments.

Apple's multi-party AI partner

Meta is just the latest AI model developer to partner with Apple.

In 2024, Apple announced a partnership with OpenAI to integrate ChatGPT into Siri, but OpenAI said that this integration “fell far short of expectations” and did not lead to OpenAI user growth. This claim comes from a document submitted by OpenAI in its XAI legal dispute with Elon Musk. Apple sued OpenAI in July, accusing it of stealing trade secrets, an accusation that OpenAI denied.

Earlier this year, Apple announced that it had reached an agreement with Google (GOOGL.US) to integrate the Gemini model into Apple Intelligence while adopting Google Cloud infrastructure.

Despite this, Apple's AI strategy has been criticized by investors and customers, in part because of its new version of Siri, which was not as capable as Head Labs' chatbot and smart agent last month.

At the same time, Meta has made every effort to bet on top AI labs, but it has been difficult to keep up with the competition, and its business still relies almost entirely on advertising. Some recent developments, such as Meta's Muse personal intelligence agent application and its latest closed-source foundation model Muse Wonder, released earlier this week, have begun to excite investors who are optimistic about Meta AI's new business opportunities.

Open source remains an important part of Zuckerberg's AI strategy. In August, Meta launched a small open weighted AI model called Muse Glimmer, and Zuckerberg promised to eventually release a larger, more powerful product.

The performance of the Muse Spark Open is not as good as the closed-source product Muse Wonder, but Wang said it is enough to attract more programmers to upgrade their hardware and try using it.

“These are not the strongest models we have developed or released,” Wang said. “However, we believe that many companies and developers will use and adapt this model on a small scale and expand the scale of deployment in the future.”