Macworld has reported on work to run Meta’s Muse Spark Open model on Apple silicon, drawing a distinction between that effort and running the Muse assistant locally. According to the publication, the configuration involves multiple high-memory Mac Studio computers, while the assistant accessed through Muse apps operates on Meta’s cloud servers.
The report attributes the model optimization work to Apple and Meta. Its central distinction concerns where processing happens: accessing an assistant through a Mac app and running model weights on Mac hardware are separate capabilities.
Muse apps and local model processing
Macworld describes Muse as an assistant accessible through Mac, iOS and web apps. In its account, those interfaces connect to an agent and models hosted by Meta; installing the app does not move that processing onto the user’s computer.
The publication identifies Muse Spark Open as a separate open-weights model covered by Apple’s optimization work. That distinction limits what the reported Mac support implies: the work concerns running that model on Apple silicon, rather than making the consumer assistant operate locally.
The reported memory requirements
Macworld puts Muse Spark Open’s parameter count at more than 800 billion and estimates that storing those parameters at BF16 precision would require approximately 1.6 terabytes of memory. The figure concerns the model weights alone, before accounting for the resources needed to run the model and operating system.
The publication reports that optimization using NVFP4, a four-bit representation, reduces the model’s memory footprint to approximately 493GB. It attributes that work to collaboration between Apple and Meta.
On Macworld’s figures, the difference between a 493GB model footprint and a computer’s total memory capacity is consequential. The report says additional runtime and macOS overhead prevent the described model from fitting into a single 512GB Mac Studio configuration.
A configuration spanning multiple Macs
Macworld describes a setup consisting of two Mac Studios, each with 512GB of memory, connected through Thunderbolt. It says Apple’s MLX framework distributes the computation across the machines.
The hardware arrangement in that account therefore extends beyond a single desktop computer. Macworld does not establish whether an alternative arrangement of three 256GB machines would work, so that possibility remains outside the configuration it describes.
The reported local-model setup also leaves a separate boundary intact: according to Macworld, running Muse Spark Open on connected Macs does not reproduce the cloud-hosted Muse assistant available through the apps.
