Apple and Meta announced a collaboration that will let developers run Muse Spark Open locally on linked Mac Studio computers. Developers will be able to use Meta's new model for coding, video analysis and other agentic tasks on their own hardware rather than in the cloud.
816 billion parameters and open weights
Muse Spark Open has 816 billion parameters and uses a mixture-of-experts architecture. Meta is releasing it with open weights, allowing developers to download the model's parameters and run it on a range of hardware. Its license permits only limited commercial use.
CNBC reports that the model is less advanced than the proprietary Muse Wonder. Meta's AI chief Alexandr Wang also said the open models are not among the most capable the company has developed or released.
An Apple silicon version and multi-computer support through MLX
Developers can download a version from services such as Hugging Face that needs no additional configuration for Apple silicon. Muse Spark Open has also been optimized to run faster on Apple's hardware.
Apple has also published technical documentation for its MLX AI framework as part of the collaboration. It explains how the framework can distribute open models across multiple computers.
Apple and Meta say running the model locally rather than processing requests in the cloud adds another layer of security.
Deployment uses two to four Mac Studio computers
Users will be able to link two to four Mac Studio computers, with the M5 Ultra chip required for this setup. A configuration with 256 GB of memory costs at least $10,799 per computer, and running the model will require multiple machines.
A Mac Studio version with 512 GB of RAM is scheduled for release later in October 2026.
