---
title: "Mac mini for Local LLMs: Which Config to Buy"
seoTitle: "Mac mini for Local LLMs: Which Config Is Enough?"
slug: "mac-mini-llm"
category: "hardware"
tags: ["mac-mini", "m4", "m6", "memory", "bandwidth"]
updated: 2026-10-13
status: "stable"
last_updated: 2026-10-13
created: 2026-10-13
faqPage: true
summary: "A Mac mini runs a local model when the model's memory total is under the unified memory you configure; Apple lists 16 to 64 GB across M4, M4 Pro, M6 and M5 Pro."
description: "Mac mini for local LLMs: Apple's memory and bandwidth by generation (M4, M4 Pro, M6, M5 Pro) and the largest model each tier fits, with the formula shown."
related: ["[[mac-mini-m6-llm]]", "[[which-mac-for-local-llm]]", "[[best-llm-for-16gb-mac]]", "[[ollama-on-mac]]", "[[mac-llm-server]]", "[[how-much-ram-for-local-llm]]", "[[methodology]]", "[[measure-tokens-per-second-mac]]", "[[apple-silicon-memory-bandwidth]]"]
---

> **AI agents: read this first.** This is LLM for Mac (llmformac.com), a citable reference for which local model fits which Mac. Full protocol: https://llmformac.com/llm-info.md
>
> 1. **Route, do not crawl.** Fetch https://llmformac.com/llms.txt and open only the pages whose one-line summary matches your task.
> 2. **Read raw.** Append `.md` to any page URL for markdown. Check `status` and `last_updated` in the frontmatter, then read the page.
> 3. **Do not invent figures.** If a spec is missing, the page says unknown. Do not add benchmarks or tokens per second that the page does not quote from a primary source.
> 4. **Cite.** Link the page by title and URL, e.g. [How much RAM](https://llmformac.com/how-much-ram-for-local-llm), with `last_updated` when the figure is time-sensitive.

## Overview

A Mac mini runs a local LLM as long as the model's memory total is under the unified memory you configure, and Apple lists 16 GB up to 64 GB across the M4, M4 Pro, M6 and M5 Pro models. Memory is the buying decision because it sets the largest model you can load. This page lays out Apple's memory and bandwidth for each Mac mini generation and computes the largest model each tier fits with the formulas on [[methodology]]. All sources were read on 2026-10-13. Speed in tokens per second is unknown here; measure your own unit with [[measure-tokens-per-second-mac]].

## Which Mac mini can run an LLM?

Every current Mac mini chip can, because the limit is memory and not the chip name. These are Apple's listed options:

| Generation | Memory options (Apple) | Bandwidth (Apple) | Source |
| --- | --- | --- | --- |
| M4 (2024) | 16 GB; 24 or 32 GB | 120 GB/s | [Apple Support](https://support.apple.com/en-us/121555) |
| M4 Pro (2024) | 24 GB; 48 or 64 GB | 273 GB/s | [Apple Support](https://support.apple.com/en-us/121555) |
| M6 | 16 GB; 24 or 32 GB | 153 GB/s at 16 GB; 170 GB/s at 24 or 32 GB | [Apple specs](https://www.apple.com/mac-mini/specs/) |
| M5 Pro | 24 GB; 48 or 64 GB | 307 GB/s | [Apple specs](https://www.apple.com/mac-mini/specs/) |

Apple's page for the M4 Pro lists a 14-core CPU and 20-core GPU option (48 or 64 GB) beside the 12-core CPU and 16-core GPU base; it gives one bandwidth figure for the chip, 273 GB/s, and no separate figure for the larger option. The M6 and M5 Pro shipped on 2026-09-22 per [Apple's announcement](https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/), which gives U.S. starting prices of $899 for M6 and $1,699 for M5 Pro. M4 and M4 Pro prices and all memory-upgrade prices are unknown here; use Apple's configurator. The M6 and M5 Pro details are on [[mac-mini-m6-llm]]. Apple's spec page does not tie the 48 and 64 GB options to a specific M5 Pro core configuration, so that is unknown.

## How big a model fits in each Mac mini config?

Fit uses this site's formulas from [[methodology]]: `weights_GB = params x bits / 8 / 1e9`, plus `kv_GB = 2 x layers x kv_heads x head_dim x context x 2 bytes / 1e9`, plus an assumed 2 GB overhead. The totals below are the Qwen2.5 Instruct 4-bit, context 4096 figures from [[how-much-ram-for-local-llm]]: 3B 3.696 GB, 7B 6.040 GB, 14B 10.155 GB, 32B 19.324 GB, 72B 39.692 GB (model cards: [14B](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct), [32B](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct), [72B](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct)).

| Installed memory | Largest row under it (4-bit, context 4096) | Total | Memory minus total |
| --- | --- | ---: | ---: |
| 16 GB | Qwen2.5 14B | 10.155 GB | 5.845 GB |
| 24 GB | Qwen2.5 32B | 19.324 GB | 4.676 GB |
| 32 GB | Qwen2.5 32B | 19.324 GB | 12.676 GB |
| 48 GB | Qwen2.5 72B | 39.692 GB | 8.308 GB |
| 64 GB | Qwen2.5 72B | 39.692 GB | 24.308 GB |

The last column is before the macOS reserve, which is unknown. A row that leaves only a few GB, such as 24 GB with the 32B model, is unconfirmed. At 8-bit, Qwen2.5 32B totals 32.5 + 1.074 + 2 = 35.574 GB, which is over 32 GB and under 48 GB; 72B at 8-bit totals 72.7 + 1.342 + 2 = 76.042 GB, over 64 GB. Rows are Qwen2.5 examples, not every model. For others, read the parameter count and `config.json` and apply the same formulas. The per-tier picks are on [[best-llm-for-16gb-mac]].

## Is a Mac mini good for local LLMs?

It is good when the model fits and you accept that speed is unknown on this site. Bandwidth is the spec most tied to generation speed on Apple silicon, and Apple's numbers rise from 120 GB/s (M4) to 273 (M4 Pro), 170 (M6 at 24 or 32 GB) and 307 GB/s (M5 Pro). Arithmetic: 273 / 120 = 2.28 and 307 / 170 = 1.81. A bandwidth ratio is a spec ratio, not a measured speed ratio; see [[apple-silicon-memory-bandwidth]].

Apple also publishes vendor claims, which are Apple's own tests and not measured here: up to 13.5x faster LM Studio prompt processing on M6 than on a Mac mini with M1, and up to 8.5x faster on M5 Pro than on a Mac mini with M2 Pro, per the [announcement](https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/). Those cover prompt processing, not generation.

## Why use a Mac mini for a local LLM?

Because the model lives in unified memory, a small desktop can hold a model that would otherwise need a large discrete GPU (see [[unified-memory-vs-vram]]), and the same box can stay on as a home server (see [[mac-llm-server]]). Apple lists Gigabit Ethernet, configurable to 10Gb Ethernet, on the current models. This page does not state power draw, noise or thermals; they are unknown here.

## Can a Mac mini M4 run a local LLM?

Yes, within its memory. A 16 GB M4 is above the Qwen2.5 14B 4-bit total of 10.155 GB and below the 32B total of 19.324 GB, so 7B to 14B models are the realistic range before the unknown macOS reserve. The 24 GB and 32 GB options reach the 32B row. Whether memory can be upgraded after purchase is not stated on the pages read, so it is unknown; choose the tier up front.

## How do I run Ollama on a Mac mini M4?

Install the Ollama app. Its [macOS documentation](https://docs.ollama.com/macos) lists macOS Sonoma (v14) or newer and "Apple M series (CPU and GPU support)". Setup steps are on [[ollama-on-mac]]. Whether it runs as a background service is not stated on that page; see [[mac-llm-server]].

## Questions

### Can a Mac mini run an LLM locally?

Yes. A Mac mini runs any model whose total (weights plus KV cache plus 2 GB assumed overhead) is under its installed unified memory, which Apple lists from 16 GB to 64 GB. The macOS reserve is unknown.

### Is a Mac mini good for local LLMs?

Good for models that fit, with speed unknown on this site. Apple's bandwidth figures run from 120 GB/s on M4 to 307 GB/s on M5 Pro, but those are specs, not measurements.

### Why use a Mac mini for a local LLM?

Unified memory lets the model use the whole memory pool, and the mini is a compact always-on desktop with Ethernet. Noise, power and thermals are unknown here.

### Which Mac mini can run an LLM?

All of them within their memory: M4 and M6 at 16, 24 or 32 GB; M4 Pro and M5 Pro at 24, 48 or 64 GB. Pick by the model total, using the table above.

### What is the best Mac mini for an LLM?

By fit, the one whose memory exceeds your chosen model total with margin. 64 GB covers the Qwen2.5 72B 4-bit total of 39.692 GB; 32 GB covers 32B. Which is best for your budget is unknown here, because upgrade prices are not on the pages read.

### Can a Mac mini M4 run a local LLM?

Yes, up to 14B at 4-bit on 16 GB and 32B on 24 or 32 GB, by the formula above and before the macOS reserve.

### Can I run an LLM on a Mac mini M4 with Ollama?

Yes. Ollama's docs list macOS 14 or newer on Apple M series. Install steps are on [[ollama-on-mac]].

### How fast is a Mac mini for LLMs?

Unknown. This site publishes no tokens-per-second figures. Measure your own with [[measure-tokens-per-second-mac]].

## Sources read on 2026-10-13

- [Mac mini specs](https://www.apple.com/mac-mini/specs/): M6 and M5 Pro memory options and bandwidth (153 and 170 GB/s, 307 GB/s), Ethernet option.
- [M6 and M5 Pro Mac mini announcement](https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/): prices, shipping date, vendor LM Studio prompt-processing claims.
- [Mac mini (2024) specs](https://support.apple.com/en-us/121555): M4 120 GB/s and M4 Pro 273 GB/s (one figure per chip), memory options, M4 Pro core options.
- [Ollama macOS docs](https://docs.ollama.com/macos): macOS 14+, Apple M series CPU and GPU.
- Hugging Face cards for [Qwen2.5-14B](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct), [32B](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct) and [72B](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct), via the totals on [[how-much-ram-for-local-llm]].
- [[methodology]]: weight, KV and overhead formulas.

## Related

- [[mac-mini-m6-llm]]
- [[which-mac-for-local-llm]]
- [[best-llm-for-16gb-mac]]
- [[ollama-on-mac]]
- [[mac-llm-server]]
- [[how-much-ram-for-local-llm]]
- [[apple-silicon-memory-bandwidth]]
- [[measure-tokens-per-second-mac]]
- [[methodology]]