The architecture is deliberately simple. A Minion Server splits submitted tasks into subtasks stored in a PostgreSQL database, and a Minion Executor on each participating Mac pulls the next subtask over REST APIs, with no central scheduler. Each Mac reads inputs from Amazon S3, fetches the model if needed, runs the subtask on its own GPU and returns the result to S3. A Minion Watchdog tracks heartbeats sent every five minutes, so unfinished subtasks return to the queue when a Mac goes to sleep, drops off the network or its owner stops the executor to reclaim the machine.

The system relies specifically on Apple Silicon and handles only AI inference, in two flavours: embeddings using open-source Hugging Face transformer models, and text generation from large language models via Ollama, installed on each participating Mac. Notably, each Mac runs a whole model on its own hardware; the paper does not suggest any single model is split across machines.

Participation is informal, often arranged through Slack messages asking colleagues to temporarily run the executor, and owners can prioritise which jobs their Mac takes first. The paper cites reduced GPU queue times, no new GPU requests and zeroed infrastructure cost, though this excludes electricity and hardware wear. Confidentiality is a real concern, with researchers noting teams worry about their data running on devices outside their group; Apple mitigates this with team- and organisation-scoped deployments using separate databases and Mac pools.

The approach echoes older volunteer-computing projects such as SETI@home and Folding@home, which likewise distributed small chunks of work to idle machines. The authors also frame the idea against a backdrop of constrained compute: Apple's July quarterly filing noted it depends on access to sufficient computing resources in a market with constrained supply, extended lead times and increasing costs.