Apple @ Work: The spiritual successor to the Xserve is here, thanks to Apple silicon and MLX - 9to5Mac

A different path in AI

For years, Apple was criticized for lacking a frontier AI model, while Siri delays and the Apple Intelligence rollout drew attention. The article argues that competing to build the largest models has become costly, as companies spend billions on GPUs and open-weight alternatives can soon run on existing hardware.

Meta’s Muse Spark Open has 816 billion parameters, 42 billion active parameters and a million-token context window. Apple did not create the model; its contribution is hardware capable of running it, though Siri remains far behind.

Mac hardware for models

Apple silicon uses unified memory, allowing the GPU in a Mac Studio to access up to 512 GB. Apple’s MLX framework is designed for that architecture and includes tools for building and optimizing machine-learning models.

A successor on the desk

The article describes two Mac Studios as able to hold an entire model, combining their memory, while the M5 Ultra offers 1.2 TB/s of memory bandwidth. Together, the machines can function like a system with a terabyte of unified memory.

Tests described in the article used Muse Code, powered by Muse Spark Open, for local workflows including changes to the iOS app Ice Cubes and video editing in DaVinci Resolve. The comparison is to Apple’s discontinued Xserve, which ended in 2011: a couple of Mac Studios can now host a model that might otherwise need rack-mounted Nvidia hardware.