Apple and Meta bring Muse Spark Open to local Mac hardware

Apple and Meta have worked together to run Meta’s 816-billion-parameter Muse Spark Open model on Mac Studio computers, according to Gadget Review. The striking part is that a model this large can be run locally, but the reported setup comes with a very large hardware bill.

Meta AI Watch analysis

What happened

Gadget Review says the collaboration uses Apple’s MLX framework to distribute Muse Spark Open across two to four M5 Ultra Mac Studios. The model has 816 billion parameters, though about 42 billion are active for each token, and a reported context window of one million tokens. Its weights are available to download, the outlet says.

The report says the local setup keeps prompts, files and outputs on the machines during inference rather than sending them to a hosted service. It also warns that the exact minimum requirements vary with quantisation and context length, so its hardware estimates are starting points, not guarantees.

Why it matters

This is a concrete route to running a frontier-scale model on local hardware, with potential appeal for developers who want to experiment without sending inference requests to a cloud provider. But “local” here does not mean “on the laptop already under your desk”: Gadget Review puts the entry cost for two base Mac Studios above $20,000.

The arrangement also gives Apple’s MLX framework a demanding use case and gives Meta another way to put its open-weight model in developers’ hands. The report describes a technical collaboration, not a Siri or Apple Intelligence integration, nor evidence of a wider partnership.

Our read

The interesting achievement is the practical demonstration, not the idea that local AI has suddenly become cheap. Keeping inference on hardware you control is appealing; buying several top-end machines to do it is a rather less casual hobby. Developers weighing the trade-off should look closely at the actual model configuration, memory needs and workload before treating the reported floor as a shopping list.

What to watch

• Whether Apple or Meta shares a supported configuration and clearer minimum requirements.
• How performance changes with quantisation and context length.
• Whether the collaboration leads to other concrete ways to run Meta models locally.

Discussion spark: Would you spend more than $20,000 to run a model of this scale locally, or is cloud inference the more sensible trade-off?

not affiliated with, endorsed by, or operated by Meta