The paper pools jobs, not memory.
Apple's Minions paper describes a central server that splits an inference task into subtasks. Executors on Apple Silicon Macs pull eligible work through REST APIs. Each Mac runs its subtask locally, then writes the result to an object store.[1]
That design is different from tensor parallelism. The Macs do not combine their memory into one address space for a single model invocation. The paper names embedding generation and language-model completion as the current workloads. Executors use Hugging Face models for embeddings and Ollama for completions.[1]
The useful unit is a job card that another machine can pick up after the first one disappears.
Device churn shapes the system.
Personal computers sleep, leave the network, and return to their owners. Minions treats those events as normal. Busy executors send heartbeats every five minutes. A watchdog releases unfinished work after a worker stops responding, so another executor can claim it.[1]
PostgreSQL holds task state. A unique constraint on the subtask identity lets competing executors race for one ownership row. One insert wins. The rest must choose another subtask. This is a small, legible lock for a fleet with no stable membership.
Queue time is the reported win.
The paper says some centralized GPU jobs previously waited more than ten hours. With the Mac pool, work often started within minutes. Its comparison table reports no new GPU requests and zero infrastructure cost.[1]
Read that last number inside the paper's accounting boundary. Existing Macs still use electricity, storage, network capacity, and operator time. The paper does not publish energy use, per-task latency, throughput, fleet size, model sizes, or a matched cost study. It reports an operational queue improvement, not a universal price or performance result.
Do not confuse two kinds of distribution.
MLX documents distributed communication backends for collective work across processes. Its Ring backend performs all-reduce and all-gather over TCP. JACCL uses Thunderbolt RDMA and is described as necessary for jobs such as tensor parallelism.[3]
Minions takes the looser route. Each worker receives a separate inference subtask. That route fits machines that can vanish. It also means the workload must split cleanly, store its inputs and outputs, tolerate a retry, and avoid depending on shared accelerator memory.
Fit the job before recruiting the fleet.
A good candidate has independent records or prompts, deterministic identity, bounded input data, and an output that can be written once or deduplicated. A poor candidate needs tight cross-device synchronization, one combined memory pool, low tail latency, or uninterrupted access to every worker.
Confidentiality is another gate. The paper says concern about data on devices outside the immediate organization limited participation. Its stated response is a team-scoped deployment with a dedicated database and isolated executor pool.[1] That reduces the group. It does not replace input classification, local cache policy, object-store permissions, deletion checks, or device-owner consent.