Apple researchers describe Minions as a system that distributes AI inference subtasks across idle Apple Silicon Macs. The paper, dated October 9, 2026, explains how PostgreSQL gives each subtask a single owner and how a watchdog returns interrupted work to the queue. The researchers’ paper describes the system’s design; Tom’s Hardware reported that Apple uses it as internal overflow capacity.
That coordination is the key idea: Minions treats available Macs as workers for separate pieces of inference work, rather than as one machine pooled together. Apple researchers report that some jobs that had faced GPU waits of more than 10 hours could often begin within minutes—a change in queue delay, not a measure of how quickly a model computes.
What Minions does with idle Macs
The paper describes two current kinds of work: generating embeddings with supported Hugging Face transformer models, and completing large language model (LLM) requests through Ollama installed on participating Macs. Embeddings are numerical representations of text that software can use to compare or retrieve information.
Minions divides submitted jobs into subtasks, then assigns eligible subtasks to Mac-based executors. The paper describes inference work distributed this way; model training is future work. The described task division assigns subtasks to Macs rather than splitting a single model’s weights across them.
How Minions assigns each inference subtask
A Minion Server validates incoming requests and splits jobs into subtasks. Executors pull eligible work through REST APIs—a standard way for software services to request data or tasks. They fetch inputs and, when needed, cache a model locally, run inference on the Apple Silicon GPU, then send results to configured object storage. The paper names AWS S3 as its main example.
PostgreSQL coordinates ownership. When an executor claims a subtask, the system records that ownership against the team, project, task name, and subtask ID. A database uniqueness constraint on those fields prevents a second ownership record for the same subtask, so two Macs do not claim the same piece of work.
By default, subtasks are handled first in, first out. Executor owners can set preferences for a team, project, or task; when preferred work is unavailable, the executor can return to the shared queue.
How Minions recovers interrupted work
Executors send a heartbeat to the system every five minutes to signal that they are still responsive. The Minion Watchdog monitors active executions. If an executor stops responding—for example, after a network loss, sleep, shutdown, or intentional termination—the watchdog releases its unfinished subtask so another executor can acquire it.
That recovery path is specific to active work: it lets the system reassign an interrupted subtask instead of leaving it tied to an unresponsive Mac.
What teams can monitor
Grafana dashboards draw on PostgreSQL data to show executor counts by status: registered, online, busy, or idle. They also show task counts by status—active, completed, or failed—and by inference type. Together, those measures give teams a view of worker availability and job outcomes.
For deployments scoped to a team or organization, the paper describes using dedicated PostgreSQL databases and separate executor pools to address confidentiality concerns.
What the queue-time result measures
Apple researchers report that some inference jobs previously waited more than 10 hours for centralized GPU access, while Minions could often let tasks begin within minutes. The comparison concerns time waiting to start. It does not report how long a task takes to finish or how many tokens or other outputs a model produces per second.
Current scope and future work
The paper describes Minions as an inference system, with embeddings and LLM completion as its current workload categories. Distributed model training is identified as future work. The authors also describe a wider pool spanning organizations or the global community as a conditional possibility, dependent on whether idle devices can be securely federated.
