Apple’s big AI week ends with Mac + Muse AI supercomputers

Apple's machine learning team released a new version of MLX that runs Meta's open-weight frontier model Muse Spark Open locally on Mac Studio with Apple silicon, with a cluster of four Mac Studios generating over 36 tokens per second, according to an Apple white paper. Meta AI chief Alexandr Wang said Apple's unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model, and the chip's 1.2 TB/s of memory bandwidth supports the per-token reads. The work, which Apple CEO John Ternus called "amazing," lets businesses run an almost trillion-parameter model on-premises on two or four Mac Studios for agentic tasks.

It is not the big news because Apple this week made huge steps in AI. Not only have its researchers come up with a nifty tool to turn Macs into distributed inferencing data centers https://www.computerworld.com/article/4235901/apples-minions-tech-turns-idle-macs-into-an-ai-data-center.html , not only does the company now have a plan for Apple Intelligence in China https://www.computerworld.com/article/4235207/china-clears-alibabas-qwen-ai-just-in-time-for-iphone-duo.html , but it has also worked with Meta https://research.meta.ai/blog/muse-spark-open to make it possible to run on-premises instances of Muse Spark Open on a couple of Mac Studios. Its machine learning team also introduced a new version of MLX https://machinelearning.apple.com/research/mlx-distributed-inference that enables AI researchers and developers to run large language models LLMs locally on Mac Studio with Apple silicon https://www.computerworld.com/article/4225059/review-m5-ultra-mac-studio-pure-unadulterated-power.html . In case you’d missed it, MLX leverages the unified memory architecture https://www.computerworld.com/article/1611155/apple-launches-mlx-machine-learning-framework-for-apple-silicon.html of Apple silicon, which is invaluable if you want to build new inference and fine-tuning methods. Over the past year, the ML team has worked on “several key innovations” to make it easier and more efficient to “run state-of-the-art LLMs locally on recent Apple silicon.” These improvements include: The release leans heavily into what Apple’s teams have been doing with Muse, and once again demonstrates how the very architecture of Apple silicon supports AI inferencing. That’s no surprise to anybody at this point: not only have I been banging on that drum https://www.computerworld.com/article/4218761/macs-dont-just-do-ai-theyre-replacing-the-cloud-for-it.html regularly, but almost anyone who is anyone in AI development and research already works on a Mac https://www.computerworld.com/article/4063294/macstadium-sees-apple-adoption-accelerating-across-us-enterprises.html . A new white paper https://machinelearning.apple.com/research/mlx-distributed-inference from Apple’s machine learning team shares what it has been doing with MLX, along with significant statistics concerning how well its systems support Muse Spark Open, Meta’s new open-weight frontier model. “On a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second, well above average human reading speed,” they said. That’s impressive. It means for the cost of four Mac Studios, a business can run a full-strength, frontier-class model with almost a trillion parameters, on-premises, securing company data while extracting whatever benefits there are to be extracted from the use of it. Apple and Meta collaborated on making it possible to run Muse Spark Open AI models on Macs. It’s a shrewd move for both companies, given the growing desire to run powerful AI models locally rather than in the cloud. As a result of this work, it’s possible to use Muse Spark Open on two or four Mac Studios to engage in all kinds of agentic tasks. Commenting on this on X, new Apple CEO John Ternus wrote https://x.com/johnternus/status/2110789142326165518 : “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 ” Meta’s AI chief Alexandr Wang pointed out that this is in part made possible by Apple’s Unified Memory architecture, which can provide sufficient memory to run these models. “Apple’s unified memory architecture lets the M5 Ultra’s integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model,” Meta explained https://research.meta.ai/blog/muse-spark-open . “And the chip’s 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs. The result is two Mac Studios behave like one supercomputer with a terabyte of unified memory.” Meta says it tested the system using agentic workflows running entirely on-device, including applying bug fixes and new features to an open-source Mastodon client and driving DaVinci Resolve to review footage and transact tasks. The use of DaVinci Resolve makes this commentator curious, given that Apple is expected to introduce new pro Macs this month. The idea that you may be able to run tasks like this on other Macs will surely be something Apple’s developers are thinking about. Step by step, Apple is defining the playground for AI https://www.computerworld.com/article/4228290/apple-wasnt-late-to-ai-it-was-the-adult-in-the-room.html . In doing so it is attempting, as Greg Joswiak, Apple’s senior VP of worldwide marketing, told TIME https://time.com/collection/best-inventions/2026/apple-iphone-duo/ , not to make AI for “the lunatic fringe,” but to create an ecosystem for artificial intelligence that will actually help people — a strategic approach that, if it comes together, will help maintain Apple’s position at the intersection of computing and the liberal arts. Read my Mac Studio review here https://www.computerworld.com/article/4225059/review-m5-ultra-mac-studio-pure-unadulterated-power.html . Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core https://thecorenews.substack.com/p/welcome-to-the-core?r=5l3lg , or follow me on BlueSky https://bsky.app/profile/jonnyevanssays.bsky.social , LinkedIn http://www.linkedin.com/in/jonnyevans , or Mastodon https://social.vivaldi.net/@jonnyevans .