Image: 9to5Mac
Why it mattersTwo Macs on a cable can run a bigger model than one Mac alone. The target is a team that wants to keep its data on its own hardware.
A Mac that runs an AI model on the machine, with no cloud call, was already possible. A Mac that runs a bigger model by sharing memory with a second Mac over a cable is new. 9to5Mac reports that Apple is testing exactly that. Ryan Christoffel at 9to5Mac wrote the story on 10 November 2026. The source is leaker Aaron Perris, who found evidence of the app in Apple's code.
The app is called Harbor and lives under Apple's Claris brand. It is in a private preview. Apple has not announced it. 9to5Mac says it is likely to launch publicly "in the weeks and months ahead", with no specific date.
What Perris says Harbor does
9to5Mac quotes the description Perris found. Harbor "turns your Mac into a private AI inference server." It runs open-weight models locally on Apple silicon. The Mac does not send data to the cloud. Users browse, download and load models from inside the app.
The part that is new is clustering. Perris says the app detects other Macs over Thunderbolt 5 and pools their unified memory as one pool. A model that would not fit in the memory of one Mac can run across two. The Mac is the inference server, and the cable turns several of them into one bigger server.
Why the hardware matters
Thunderbolt 5 came with the M4 Mac family in late 2024. It runs at 80 Gbps in each direction. A one-way 120 Gbps mode is available for displays. That is fast enough. A tensor crossing the cable finishes in microseconds, so the cable itself does not slow a model down. Apple silicon uses unified memory, shared by the CPU and the GPU. The pooling is built on that shared memory. A model's weights live in one pool and every Mac in the cluster reads from the same addresses.
Nothing on an Apple silicon Mac today stops a single machine from running local models. What stops a team is the size of the memory. A Mac mini tops out well below a model a team may want to run in production. Two Macs on a cable change that answer.
A third option for private deployment
A team that cannot send customer data to a hosted model has been told to run its own. The usual answer is a GPU server or a cloud GPU subscription. A Mac with another Mac bolted to it is a third option, if the clustering works outside the lab.
Nothing about Harbor is confirmed yet. The story is a leak and the launch has no date. The signal is what to pay attention to today. Apple is treating local inference on Apple silicon as a real product it plans to release. Running a model from an office closet looks like part of its roadmap. Apple has not said much about it in public. A team that writes software for regulated customers now has a reason to track this one.
Source
Apple has new Mac app coming to make running local AI models easy, by Ryan Christoffel at 9to5Mac.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
