Apple’s Minions pools idle Macs to cut AI inference waits from 10 hours to minutes
Apple researchers describe Minions, an internal system that pools the GPUs of idle Apple silicon Macs to run machine-learning inference tasks. They report that jobs which previously waited more than 10 hours for central GPU capacity can often start within minutes; Apple has not announced plans to make the system available outside the company.
Bottom line — Minions currently handles inference, with distributed model training listed as future work.
Go deeper 6
-
Apple’s paper says Minions complements central GPU infrastructure rather than replacing it.
-
The system supports embedding generation and large language model completions, using models from Hugging Face and Ollama.
-
Apple says Ollama runs models locally on participating Macs, without remote API calls.
-
A watchdog returns a task to the queue if a Mac sleeps, goes offline or is taken out of service.
-
Apple’s researchers say deployments can be scoped to a team or organisation to address data confidentiality concerns.
-
The paper reports Minions has supported inference workloads across multiple Apple teams.