Mac Studio and MLX emerge as the spiritual successor to the Xserve

What happened

Apple discontinued the Xserve rack-mounted server in 2011, but a 9to5 Mac column argues its spiritual successor now sits on a desk rather than in a data centre. The claim rests on Apple silicon's unified memory and the MLX machine learning framework, which together let Macs run very large AI models locally.

The column, written by long-serving Apple IT administrator Bradley Chambers, points to two recent developments. Meta optimised its most capable open-weight model to run on a Mac Studio, and the open-source Osaurus app runs usable local AI on a 16 GB Apple silicon laptop.

Why Mac hardware suits large models

Apple silicon uses unified memory, meaning the processor and graphics share one pool of RAM. On a Mac Studio, the integrated GPU can address up to 512 GB of that memory, far more than a typical graphics card holds.

According to Apple's Machine Learning Research post cited by 9to5 Mac, four M5 Ultra Mac Studios running Meta's Muse Spark Open model exceeded 36 tokens per second, while each machine drew roughly 300 watts at peak. The M5 Ultra's memory bandwidth of 1.2 TB/s is fast enough to serve the parts of the model each token needs. Two Mac Studios can therefore hold an entire model between them, behaving like one machine with a terabyte of unified memory.

MLX and real-world workloads

MLX is Apple's machine learning framework, built for the unified memory architecture of Apple silicon. Its interface resembles NumPy, and it includes packages for neural networks, optimisers, automatic differentiation and graph optimisation. Bindings for Swift, C++ and C let it run across Apple platforms.

The column describes agentic workflows running entirely on-device, using the Muse Code tool powered by Muse Spark Open. These included shipping real iOS changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve to review footage frame by frame and crop the strongest moments vertically.

Local AI on ordinary MacBooks

For IT teams, the smaller end of the lineup may matter most. Osaurus is a free, open-source Mac app that manages local models which work offline, and can call a frontier model when needed. Before sending anything to Claude or OpenAI, it runs a small local model to find names, phone numbers and account numbers, replacing them with placeholders.

Osaurus chief executive and co-founder Terence Pae told the Apple @ Work podcast that the company keeps no servers and no customer database, so the app would keep running even if the business closed. Individuals use it free, while a paid tier called Workspaces lets teams share agents across their own Macs. A Windows beta also exists.

Why it matters

The column argues Apple chose not to compete with OpenAI, Anthropic or Nvidia in frontier models, and instead built the memory, bandwidth and software that run other people's models. Meta's Muse Spark Open is described as the most capable open-weight model in the United States, and it is now optimised for the Mac.

The practical result is that a couple of quiet Mac Studios under a desk can host a model that would otherwise need a rack of Nvidia hardware and a matching power budget. Siri, however, remains far behind, and the column's claims about workflows come from its author's own testing, so readers should treat them as reported rather than independently verified.