Running an open AI model on your own Mac usually means installing a third-party app and learning which model fits in how much memory. A leak suggests Apple wants to make that much simpler. According to 9to5Mac, the leaker Aaron Perris has found evidence of an Apple app called Harbor that turns a Mac into a private AI inference server, and can link several Macs together to run models too big for one machine.
The description Perris posted is specific. Harbor runs open-weight models locally on Apple silicon, and the text stresses that your data never leaves your hardware and nothing reaches the cloud. You browse, download and load models from inside the app, and Harbor handles model management. The most interesting feature is clustering: connect two or more Macs with Thunderbolt 5 and Harbor pools their unified memory into a single resource, either to run a model larger than any one computer can hold or to serve more requests at once. Setup is described as automatic, with no manual configuration or command line. It is reportedly in private preview under Claris, Apple's software subsidiary.
That would put Apple in direct competition with tools we review. LM Studio and Ollama already let Mac users download and run open models with a few clicks or one command, and both offer a local server that other apps can call. Much of the speed on Macs comes from MLX, Apple's own open-source machine learning framework, which many of these apps use. What Harbor would add is first-party polish and, above all, easy multi-Mac pooling, which today takes hobbyist set-ups and some patience.
The timing fits. 9to5Mac notes that Macs have grown more popular this year, partly as machines for running models locally on a Mac Studio or Mac mini. A Mac with a large pool of unified memory is one of the few consumer machines that can hold a big model, and linking two or three of them is a cheaper route to more memory than a rack of data-centre GPUs.
The caveats are those of any leak. Apple has not announced Harbor, the description comes from a single post, and private previews do not always ship. Nothing is known about pricing, which Macs will be supported, which models it will offer or how fast clustered inference will be in practice. Thunderbolt 5 also limits clustering to recent, higher-end Macs.