No Apple document names Harbor; 9to5Mac’s Nov. 10, 2026 item cites one leaker post

9to5Mac’s November 10, 2026 item cites no Apple press release, developer note, or App Store listing. It cites an X post by Aaron Perris, @aaronp613, describing Harbor as a private-preview app under Apple’s Claris brand. The described behavior is on-device open-weight inference on Apple silicon, no cloud egress, and unified-memory pooling across Thunderbolt 5 Macs. The article supplies no model list, build number, throughput figure, or pooling protocol. A leaker post is not a shipping record.

Checkable substrate is not Harbor. Apple’s developer documentation describes unified memory on Apple silicon. MLX, Apple’s array framework for that hardware, is public on GitHub under ml-explore. Thunderbolt 5 is a published interconnect at up to 80 Gbit/s bidirectional; Apple’s M4 Pro and M4 Max Mac specifications list Thunderbolt 5 on specific ports. None of those documents name Harbor, a Claris inference cluster, or a cross-Mac memory API. Feature bullets in a leak are claims, not measurements.

The coverage equates a setup wizard with a new inference capability. Pooling unified memory over Thunderbolt is not on-package NUMA. Latency, bandwidth, and failure domain change, and only a protocol spec or benchmark can price that gap. llama.cpp with Metal, Ollama, and LM Studio already run open-weight models on Apple silicon. A GUI does not alter weights, quantization, or tokens per second. No document supports a claim that Harbor changes daily workflows inside 12 months.

Operational status is unchanged until Apple ships a binary, a support note, or a measurable endpoint. The test is an apple.com or claris.com URL that names Harbor, lists supported models, and states cluster limits. Private preview, if the Claris attribution is accurate, still requires a distribution path: TestFlight, a Claris portal, or an AppleSeed build. Absent that artifact, Harbor remains a string in a secondary report. Do not budget hardware or workflow changes on it.