Author: Merrel Briona

Apple has been revealed to be developing “Harbor,” an app for running open-weight AI models locally on Macs. Its hallmark design connects multiple Macs via Thunderbolt 5, pooling their unified memory to power inference of large-scale models, completing processing without sending data to the cloud. Development is handled by Apple subsidiary Claris.

The first to flag Harbor’s existence was Aaron Perris, who is well versed in Apple-related information. On November 10, he posted on X screenshots of the app apparently obtained from an investigation of system files. Claris’ website hosts a Harbor page, which states that it can run open-weight models on Apple Silicon Macs, that setup takes only minutes, that multiple Macs can be connected via Thunderbolt 5 to share memory, and that data is stored on the Mac and not exposed externally.

At the core of Harbor is a feature that bundles multiple Macs into a single inference resource. The constraint when running large language models is memory capacity rather than compute power, and models too large for a single machine can be run by combining the unified memory of multiple machines. Perris’ post includes screenshots of an M5-chip MacBook Pro running “Llama 3.6-2-Instruct-4bit” and of three Macs connected via Thunderbolt 5 and managed as a single cluster. It also shows the app downloading an MLX-compatible version of “gpt-oss-20b.”

The feature description repeatedly emphasizes ease of introduction. Devices are automatically detected via Thunderbolt 5, and manual configuration, command-line operations, and infrastructure expertise are said to be unnecessary. Model search, download, and loading are all completed within the app, designed so users can focus on the work itself rather than environment setup. The intended users are not MLOps specialists but rather those who want to run models in their own creative environment without delving into CUDA or vLLM configuration.

Behind Apple’s move into this area is a shift in the structure of Mac demand. Buying Mac Studio or Mac mini machines for local model execution is spreading in the community, and there is a view that this, alongside inexpensive models such as the MacBook Neo, is a factor pushing up Mac sales. Currently, however, building a local model environment on a Mac requires relying on third-party tools such as Ollama, LM Studio, and llama.cpp, and clustering requires MPI configuration or the use of projects like exo. If Harbor achieves a “just connect and use” level, it would fill a gap in Apple’s own product lineup.

Harbor is currently in private preview and is placed under the Claris brand. Claris is a wholly owned Apple subsidiary that has handled productivity software such as FileMaker. The placement suggests the project may have emerged from the productivity-tools lineage rather than from Siri or the machine learning division. Apple has not officially confirmed Harbor and has not announced a formal release date.

In the same vein, Apple is also developing “Minions,” an app that connects multiple Macs to build a virtual GPU cloud. The concept aggregates idle Macs to dramatically shorten AI processing latency, and it may share a technical foundation with Harbor’s clustering feature.

All information obtained so far is based on Perris’ analysis of system files and is not an official Apple announcement. It cannot be determined whether the confirmed feature descriptions are actual marketing copy or internal drafts, and there is no guarantee that the final product specifications will match them.

The market impact lies in expanding the Mac’s positioning from “a device for using AI” to “an execution platform for AI.” In particular, it could serve as material underpinning demand for higher-priced Mac Studio models and large-memory configurations. For enterprises operating models without exposing confidential data externally, it adds an alternative to cloud APIs. However, neither an announcement nor a timeline has been finalized, and whether it will replace the ecosystem built by third-party tools depends on the number of supported models and usability at general availability.