Apple described how to run Muse Spark Open locally on four Mac Studio computers
Apple presented a procedure for running Muse Spark Open locally using MLX and RDMA. According to its own measurements, it achieved a speed of over 36 tokens per second on four Mac Studio computers. The model has 816 billion parameters, of which 42 billion are active.
Apple described how to run Muse Spark Open locally across multiple computers using the MLX framework. According to its own measurements, it achieved a generation speed of over 36 tokens per second on a setup of four Mac Studio computers. The model uses a mixture of experts architecture: of its total 816 billion parameters, 42 billion are active during computation. The company evaluated it at NVFP4 precision.
According to Apple, the model parameters alone require more than 1.6 TB of memory at the original BF16 precision. Four-bit quantization to NVFP4 reduces the memory footprint to approximately 493 GB. Once intermediate computations are accounted for, however, the model does not fit even on a single Mac Studio with 512 GB of unified memory and therefore needs to be split across multiple machines.
The article explains, for example, how to split the model layers between two computers: the first processes its portion and passes the result to the second. Communication between the machines uses RDMA over Thunderbolt 5; macOS 26.2 introduced the relevant driver, and MLX has included support for it since version 0.30.1. Details are available in the source article.
Why it matters
The procedure provides a concrete basis for assessing local experiments with a large model in a private environment. Planning the setup requires accounting for intermediate computations and connections between computers, as well as memory for the parameters.
Two audiences, two different impacts
What this means
For individuals
The described procedure gives developers a basis for assessing the feasibility of their own local experiments with Muse Spark Open.
For a business
For enterprise development, this configuration requires planning the capacity of multiple computers and their connections; the memory footprint of the quantized model alone is insufficient to size the setup.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.