Global Tech News For people figuring out what to build with AI.
AI

Apple Reports 36.73 Tokens per Second Running Muse Spark Open on Four Mac Studios

A Mac Studio and Studio Display on display in an Apple Store; not the computers used in Apple's distributed-inference test.

Apple says Meta’s Muse Spark Open does not fit on one Mac Studio with 512GB of unified memory: its weights alone occupy about 493GB, before the extra memory needed to run the model. Apple therefore split it across four M5 Ultra Mac Studios. The four computers generated 36.73 tokens per second in an MLX benchmark with an 8,192-token prompt; a token is a small unit of text. NVFP4 is a four-bit format that compresses model weights to reduce their memory footprint. This was Apple’s own benchmark, not an independent comparison across computer brands.

Four computers share the inference work

Meta’s model card lists 816 billion parameters overall and 42 billion active for each token. The full weight set still needs to be available to the system. Apple says the model’s NVFP4 weights occupy about 493GB, and the additional memory needed during inference means the model does not fit on one Mac Studio with 512GB of unified memory. Apple split it across several Mac Studios connected through Thunderbolt 5 and Remote Direct Memory Access, or RDMA.

Apple tested two ways to divide the work. Pipeline parallelization assigns successive groups of model layers to different computers. Tensor parallelization divides the calculations within each layer among the computers. For Muse Spark Open, Apple reports 15.74 tokens per second with two computers using pipeline parallelization, 24.99 with two using tensor parallelization, and 36.73 with four using tensor parallelization.

Apple’s rate is tied to one benchmark workload

The benchmark used an 8,192-token prompt, generated 128 tokens and ran with batch size 1. Its tokens-per-second figures measure generation after the prompt, not end-to-end task time. The article reports Apple’s own tests and offers no independent replication or comparison with hosted services and other manufacturers’ hardware; the figures also do not measure answer quality.

Related coverage: Meta says Muse Spark Open needs 700GB to 2TB of GPU memory locally.

Image: Photo by AzureSaturn via Wikimedia Commons, CC0 1.0. Illustrative image of one Mac Studio on display in an Apple Store, not Apple’s four-computer test setup. License details.

Sources