Meta's Muse Spark Open Model Runs on Two Mac Studios, Not a Data Center

Meta released Muse Spark Open today, an open-weight AI model it calls the most capable of its kind to ship out of the United States. The company's more concrete claim is about where it runs: Meta says the model is the first in its class that can operate on a desktop machine, specifically two Apple M5 Ultra Mac Studios running Apple's MLX framework. That pairing puts a large open model on hardware a well-funded studio or research lab could buy outright rather than rent by the hour.

The two-Mac requirement is the number that matters. Apple's M5 Ultra Mac Studio tops out with substantial unified memory, and MLX lets developers spread a model's weights and computation across multiple machines over a fast local connection. Meta's framing suggests Muse Spark Open was tuned for that memory architecture rather than squeezed onto a single consumer GPU, which is where most open-weight models of this size end up running.

Apple's involvement is the second half of the story. MLX is Apple's own array framework for silicon, and Meta naming it explicitly as the runtime signals coordination between the two companies rather than a model that merely happens to compile on Macs. The Verge reported the arrangement as a Meta-Apple collaboration. Neither company has published throughput, latency or context-length figures for the two-machine setup, so the practical speed of the configuration is still unverified.

Why this lands now: open-weight releases have been drifting toward larger parameter counts that assume cloud inference or multi-GPU servers. A desktop path changes the audience. Researchers who cannot get cluster time, small teams with privacy constraints, and anyone working with data that cannot leave the building get an option that keeps weights and prompts on local hardware, provided they can afford the two machines.

The comparison Meta invites is with other US open-weight releases, and the company is claiming the top spot among them. That claim will be tested quickly by independent evaluations and by developers who try to reproduce the two-Mac setup on their own benches. Memory bandwidth, not raw compute, usually decides whether a configuration like this is usable or merely possible.

What to watch is whether Meta publishes benchmarks for the local configuration specifically, and whether MLX adds support for spanning more than two Macs. If it does, a rack of Mac Studios becomes a plausible alternative to a rented GPU node for a narrow but real set of workloads. The open question is cost: two fully loaded M5 Ultra Mac Studios are not cheap, and Meta has not said what the setup is expected to replace.