Meta and Apple team up to make Muse AI models run better on Macs

The Verge reports Meta launched Muse Spark Open as what it calls the most capable U.S. open-weight model to date, and the first in its class that can run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework. Mark Zuckerberg posted that the weights release runs great on new Mac Studios thanks to work with Apple’s John Ternus; Ternus replied congratulating the team. Meta chief AI officer Alexandr Wang told Axios the companies have collaborated for several months.

Meta describes the model as natively multimodal with 816 billion total parameters, strong on long-horizon tasks, complex software engineering, and multi-step reasoning. Apple also published a technical post on MLX covering Muse Spark Open and two other open models.

Open weights that actually fit a desk — even a very expensive desk — change who gets to run frontier-class local models. Two M5 Ultra Mac Studios is not a hobbyist build, but it’s a different story than “rent a cluster.”

The politics are almost as interesting as the silicon. Meta and Apple have spent years as platform antagonists; Wang’s “mommy and daddy playing nice” line is joking, but the months of MLX collaboration are not. Apple gets a showcase for Mac Studio as an AI workstation. Meta gets a credible local-run story next to its API and cloud partners.

Useful AI wins when serious models leave the hyperscaler cage. The downside is familiar: local doesn’t mean harmless, and an 816B-class MoE still concentrates capability in whoever can afford the RAM. I’m watching how far MLX tooling goes beyond the launch demo, and whether “runs on Mac Studio” becomes a real developer path or a keynote flex.

TechCrunch separately notes Engadget’s report that running Muse Spark Open locally still requires at least 1TB of RAM, with best results on M5 Ultra Mac Studio.