Apple's Claris Previews Harbor: A Private AI Inference Server for Apple Silicon Macs That Runs Open-Weight Models Locally and Pools Unified Memory Across Thunderbolt 5 Clusters, Early-Access Sign-Up Only

Background and the problem

Many companies want to use large models but cannot send data off their own network, and per-token cloud APIs get expensive and unpredictable under heavy use. Running open-weight models on a Mac is not new: Ollama, LM Studio and MLX already exist. But clustering machines, managing models and serving them still takes command-line and ops skills. Harbor is the first time Apple's own ecosystem tackles this as an official product, through its business-focused subsidiary Claris.

Architecture and how it works

According to the official Claris page, Harbor turns Apple silicon computers you own into a private AI server. One Mac can host an endpoint in minutes, and for larger models several machines connect over Thunderbolt 5 to pool their unified memory into one inference resource. The pulled App Store text, quoted by MacRumors, adds automatic discovery of Thunderbolt 5 devices to form a cluster, in-app model browsing and downloads, and no command line. MacRumors notes that macOS Tahoe 26.2 added RDMA over Thunderbolt 5, which speeds up clustering.

Benchmarks

Claris has not published any throughput, latency or maximum model size figures, nor a list of supported models. Leaked screenshots and copy describe the product's positioning only; real performance will have to wait for hands-on preview builds.

Getting started

For now the only step is the sign-up form on the Claris page, which asks for a company name and so appears aimed at businesses. If you plan to cluster, check Apple's "Identify the ports on your Mac" page first, since not every recent Mac has Thunderbolt 5. Until Harbor ships, Ollama, LM Studio or MLX remain the ready-made options for local inference on a Mac.