Meta's Muse Runs on Mac Studio, Giving Apple a Serious AI Showcase

Meta and Apple have produced a concrete, if narrow, example of collaboration: Meta's Muse Spark Open model can run locally on a cluster of two to four Mac Studio computers powered by Apple's M5 Ultra chip. Meta released the model's weights on October 15 after months of technical work with Apple.

Muse Spark Open is a mixture-of-experts model with 816 billion total parameters, but it activates roughly 42 billion for each token. In simple terms, only part of the model works on each piece of text, reducing the computing and memory burden compared with activating the entire model at once. Meta says the design brings capable agent-style workloads to desktop hardware.

The setup is still aimed at researchers and professional developers, not ordinary Mac buyers. A Mac Studio with an M5 Ultra and 256 gigabytes of memory starts at $10,799, and running multiple machines raises the cost substantially. Meta's documentation also says the model passed the release thresholds in its safety framework, but open weights still shift deployment and security decisions to the operator.

Apple gets an unusually demanding demonstration of its chips, unified memory and MLX software framework. Meta gets another route to developers without paying the inference bill for every user. Local operation can be attractive when teams want sensitive data to stay on their own hardware or want predictable computing costs.

The companies did not disclose financial terms, exclusivity or a duration for the work. The Financial Times described the launch as a rare collaboration between long-time rivals, but the evidence does not support assuming that it extends to advertising, Siri, smart glasses or other products. Muse can also run on non-Apple systems, and Meta has not made its most capable closed models available in the same form.

The investment case is therefore about ecosystem reach, not an immediate revenue event. Apple can point to a large model running on premium Macs, while Meta can broaden use of its open model family. The next meaningful test will be whether more large models receive first-class MLX support and whether developers build durable tools around them, rather than whether executives exchange public compliments.