Executable programs and extensions

What is Apple Minions? Sharing inference jobs across idle Macs

Give independent document jobs to different Macs.

Minions is a distributed inference¹ system published by Apple researchers. It sends embedding jobs or LLM² requests to the GPUs of multiple Apple Silicon Macs, then collects their results. Teams processing document collections can consider sharing the work across Macs when they are idle.

Building document search requires time to turn each document into a retrieval vector. When many such jobs are waiting, consider distributing batches to available Macs. This is the use case in Apple's October 2026 research. Apple Minions

People's Macs become available at different times. In Minions, an available Mac pulls a waiting job. When a device sleeps or loses its connection, its unfinished job returns to the queue for another Mac to take.

Sharing document embeddings across Macs

Embedding turns a document's content into a numerical vector for search. Use the same model for every document, and you can later retrieve documents similar to a question. Since batches can be processed independently, different Macs can work on different documents.

Before building document search, agree on document IDs, the model version, and how documents are split. All Macs need to produce vectors using the same rules for comparison in one search index. Keep document IDs with the results so you can identify and remove duplicates after a retry.

A desk with a compact desktop and PC tower beside a monitor — Each Mac must be able to run the model assigned to its documents.

When to use a job queue or model sharding

The published design uses Hugging Face Transformers for embeddings and Ollama on each Mac for LLM completion. An executor caches the required model and computes the jobs it claims. Research paper

A model that does not fit in one Mac's memory needs a runtime³ with model sharding. If each Mac can run the model and you have many documents, you can distribute the documents instead. The design depends on whether you need a larger model or more documents processed by the same model.

Ways to use multiple Macs
Your goal	Setup	First check
Process a document collection	Job-level scheduling	Model fit on each Mac
Run a model too large for one Mac	Runtime with model sharding	Runtime and inter-device communication

Decide where documents may be sent

The published design stores inputs and results in S3 and manages job status with a server and PostgreSQL. Each organization can designate its own participating Macs. For internal documents, choose the devices and restrict access to storage and job records to the people responsible for that work.

Start a personal multi-Mac experiment with public documents. Decide where originals are copied and when results and logs are deleted before extending it to sensitive documents. Manage both the machines performing inference and the route used to deliver files.

A workspace with two compact desktops connected by cable — Choose independent document jobs or model sharding for the capacity you need.

How to judge whether your workload benefits

For daily batch processing, record when the entire collection finishes rather than just one document's response time. Separate model download, queue wait, inference, and result collection to see where time was saved. Include failed and retried documents when assessing throughput⁴.

Keeping several Macs running also increases electricity use. Enter each Mac's active hours and power in the electricity calculator, then compare total completion time and monthly cost with your existing server. If one machine already meets the deadline, weigh the management overhead first.

Mac runtime and electricity cost calculator

A meeting table with a laptop, microphone, and notebooks — Include data transfer and retries in the total completion time.

Terminology notes

Inference — The process of using a trained model to compute an output for an input. Here, local inference means running the model on the user’s device.
Back to the text
Large language model — A model trained on large text datasets to learn language patterns and process or generate text in context.
Back to the text
Runtime — The software environment that provides facilities needed while a program runs. In local AI, it can also refer to a model execution engine.
Back to the text
Throughput — The amount of work processed or generated over time. Comparisons need the unit, such as tokens per second or requests per second.
Back to the text

Frequently asked questions

What should you check first in this article?: Minions is a distributed inference system published by Apple researchers. It sends embedding jobs or LLM requests to the GPUs of multiple Apple Silicon Macs, then collects their results. Teams processing document collections can consider sharing the work across Macs when they are idle.
How much memory does it need?: A model that does not fit in one Mac's memory needs a runtime with model sharding. If each Mac can run the model and you have many documents, you can distribute the documents instead. The design depends on whether you need a larger model or more documents processed by the same model.