Apple and Meta show Muse Spark Open running across Mac Studios

Apple and Meta have outlined a way to run Meta’s 816-billion-parameter Muse Spark Open model locally across Apple silicon, turning multiple Mac Studios into a surprisingly serious desktop AI setup.¹²

Meta says Muse Spark Open is an open-weight, natively multimodal model designed for coding, agentic work and long-horizon tasks. The company says it has released the weights through several cloud and model platforms, while also optimising the model to run on Apple’s M5 Ultra Mac Studio using the MLX framework.²

Apple’s own machine-learning team has now detailed the mechanics. Its MLX update supports distributed inference across several Macs, using low-latency RDMA over Thunderbolt 5. Muse Spark Open has 816 billion total parameters, with 42 billion active in its mixture-of-experts design. In Apple’s testing, four Mac Studios generated more than 36 tokens per second, while the model can be split across two machines in a pipeline configuration.¹

The hardware requirement is the catch. Apple says the model needs more than 1.6TB of memory in its native format. A 4-bit version reduces that to about 493GB before intermediate computation is counted, which means a single 512GB Mac Studio is not enough for a complete practical setup. Two M5 Ultra Mac Studios can hold the model between them, while four machines delivered Apple’s headline benchmark.¹

Our opinion

This is a meaningful demonstration of local AI, but it is not a secret way to turn one Mac mini into a data centre. The story is really about memory bandwidth, inter-machine links and software finally making expensive hardware behave like a coherent system. For researchers and developers who already need large local models, that is compelling. For everyone else, the price of the multi-Mac setup remains the obvious barrier.

Apple and Meta’s work nevertheless points to a useful direction: powerful models do not always have to live in a remote cloud. MLX, RDMA over Thunderbolt and open model weights make that possibility more tangible, even if Muse Spark Open is still very much an enthusiast and professional workload.¹²