By Rebecca Stone

Four Mac Studios now run Meta's Muse Spark Open at over 36 tokens per second, Apple's machine learning team disclosed this week. The figure, published in a new white paper on MLX distributed inference, anchors a cluster of announcements that move frontier-class AI inferencing from cloud data centers onto on-premises Apple hardware.

Apple also introduced a new version of MLX, its machine learning framework for Apple silicon, and confirmed a plan to bring Apple Intelligence to China through Alibaba's cleared Qwen AI. The company released research on turning fleets of idle Macs into distributed inferencing data centers. Friday also marked pre-order day for the iPhone Duo, though the AI announcements carried the week's strategic weight.

What is new in MLX?

MLX exploits the unified memory architecture of Apple silicon, the design in which CPU and GPU share a single memory pool. Over the past year Apple's machine learning team delivered what it called "several key innovations" to make it easier and more efficient to "run state-of-the-art LLMs locally on recent Apple silicon." The improvements include:

• Software support for the Neural Accelerators in the GPU of the M5 chip family
• Low-latency communication across several Macs using Remote Direct Memory Access (RDMA) over Thunderbolt
• Efficient implementations of recent architectures such as mixture of experts (MoE), a design that activates only part of a model per query

The release draws directly on Apple's work with Meta to run Muse Spark Open, Meta's new open-weight frontier model, on clusters of Mac Studio units.

What does the 36 tokens per second figure mean?

Apple's white paper reports measured throughput for Muse Spark Open, a frontier-class model with almost a trillion parameters, on a four-machine Mac Studio cluster. "On a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second, well above average human reading speed," the team said.

The practical consequence: for the cost of four Mac Studios, an enterprise can operate a full-strength frontier model entirely on-premises, keeping company data inside its own perimeter.

How do two Mac Studios act as one supercomputer?

The two companies collaborated on the deployment, and Meta documented the mechanics. "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. "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's AI chief Alexandr Wang credited the unified memory architecture with supplying sufficient memory to run models of this size. Apple CEO John Ternus, writing on X, said: "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 reported it tested the configuration with 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 execute tasks. Muse Spark Open runs on two or four Mac Studios for agentic workloads, according to the companies.

Why does the architecture matter for buyers?

The announcements position Apple hardware as a substitute for cloud AI inferencing, not merely a client for it. The M5 Ultra Mac Studio, reviewed by Computerworld as pure processing power, provides the memory capacity and bandwidth that frontier models require, and MLX's new RDMA-over-Thunderbolt support extends clusters across multiple machines without data-center networking.

Apple's strategy, as senior VP of worldwide marketing Greg Joswiak told TIME, is not to build AI for "the lunatic fringe" but to create an ecosystem that helps people. Apple is expected to introduce new pro Macs this month, which would extend the range of hardware capable of these workloads.