---
title: "Apple’s 20B Local Model: Test the Small Jobs First"
description: "A new single-Mac study shows a small memory footprint with important limits. Start with extraction or tagging and measure your own workload."
date: "2026-10-26"
tags: ["models", "workflow", "developer-tools"]
canonical: "https://news.maestromojo.com/news/apple-local-model-memory-small-task-test/"
---

**Maestro’s take.** Try Apple’s built-in model on a small extraction or tagging job. Its memory design is interesting. Judge it by correct answers and the memory your app actually needs.

## TL;DR

In an [October 25 practitioner study](https://www.airealist.ai/p/exploring-apple-on-device-model), Julien Simon measured Core Advanced on one M3 Max MacBook Pro with 48 GB of memory and macOS 27.0.1.

His [published data](https://github.com/juliensimon/apple-fm-mac) reports about 2.0 GiB for the model and safety processes two seconds after a 5,000-token request. A snapshot during the request was about 3.4 GiB. The 2.0 figure is not a peak-memory guarantee. The measurement excludes the operating system’s evictable file cache.

[Apple describes](https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models) a 20-billion-parameter model that activates 1–4 billion parameters at a time. It keeps the full model in flash and loads selected experts into memory. This is an existing model; Simon’s measurements are the new evidence.

## Why Maestro users care

A local model can earn its place by handling a narrow job while leaving room for the rest of your app. A large parameter count alone does not establish coding ability.

[Apple’s developer guide](https://developer.apple.com/documentation/foundationmodels/generating-content-and-performing-tasks-with-foundation-models) recommends tasks such as extraction and classification. It lists code creation, basic math and logical reasoning among capabilities to avoid.

Maestro would test this as a component for short tasks. One machine’s measurements do not establish performance across supported Macs.

## One thing to try

On an eligible device, check model availability first. Pick 30 representative tagging or extraction examples with known answers. Measure correctness, response time, and memory during requests. Keep a fallback for unavailable models or failed answers.

*Published October 26, 2026 · Practitioner study and Maestro commentary · Tags: models, workflow, developer tools*

**Original evidence:** [Simon’s study](https://www.airealist.ai/p/exploring-apple-on-device-model) and [measurement repository](https://github.com/juliensimon/apple-fm-mac). **Technical references:** [Apple’s architecture](https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models) and [developer task guidance](https://developer.apple.com/documentation/foundationmodels/generating-content-and-performing-tasks-with-foundation-models).