Guide
Local AI models on a Mac mini: what runs on each memory size
Which open models fit on each Mac mini memory size, how fast they run, and how to use a local model next to a cloud model in an AI agent. Updated for September 2026.
Updated 2026-09-15
A Mac mini can run language models on its own, with no cloud and no cost per request. What it can run depends almost entirely on one number: its memory.
This guide shows what fits on each Mac mini, what it is good for, and how to use a local model inside an AI agent.
Why a Mac mini is good at this
Model names checked in September 2026. New models appear every few months, but the memory rules below stay the same.
On most computers, a language model has to fit in the graphics card's memory, which is small and expensive. Apple silicon shares one pool of memory between the processor and the graphics cores. A Mac mini with 48 GB can load a model that would need a large, dedicated graphics card elsewhere.
Two numbers matter:
- Memory size decides which models fit at all.
- Memory bandwidth decides how fast they answer. It is much higher on Pro chips than on base chips, and higher on each new generation.
How much memory a model needs
Models are usually run quantized: their numbers are stored with fewer bits, which cuts the size with only a small loss in quality. 4-bit is the common choice.
A useful rule: at 4-bit, a model needs about 0.6 GB per billion parameters, plus room for the conversation (the context) and for macOS itself.
Where the memory goes, at 4-bit
- 16 GBA 9B model, about 6 GB
- 24 GBA 24B model, about 15 GB
- 48 GBA 35B model, about 21 GB
- 64 GBA 70B model, about 43 GB
The rest is macOS and the conversation. A model that fits with a short chat may not fit with a long document in it.
| Model size | Memory at 4-bit, roughly |
|---|---|
| 4B | 3 GB |
| 9B | 6 GB |
| 20 to 24B | 13 to 15 GB |
| 31 to 35B | 18 to 21 GB |
| 70B | 43 GB |
Two things change the picture:
- Long context uses memory too. A model that fits with a short conversation may not fit with a 100-page document in it.
- macOS gives the graphics cores a large share of memory, but not all of it. Leave headroom, especially on 16 and 24 GB machines.
What runs on each Mac mini
| Memory | Comfortable model sizes | Examples (September 2026) | Good for |
|---|---|---|---|
| 8 GB | Up to about 4B | Gemma 4 E2B, E4B | Sorting and classifying messages, extracting fields from text |
| 16 GB | Up to about 9B | Qwen3.5 9B, Gemma 4 E4B | Summaries, simple agents, routing requests, short drafts |
| 24 GB | Up to about 24B | Mistral Small 3.2 24B, gpt-oss-20b | Tool calling, coding help, document questions |
| 32 GB | Up to about 35B | Gemma 4 31B, Qwen3.6 35B-A3B | Stronger reasoning, writing, agent work with images |
| 48 GB | 35B with long context, or several models at once | Gemma 4 31B, Qwen3.6 35B-A3B | A serious local assistant next to other workloads |
| 64 GB | Up to about 70B | Llama 3.3 70B | The largest models a Mac mini can hold |
About "A3B" models. Names such as Qwen3.6 35B-A3B mean a mixture-of-experts model: 35 billion parameters stored, but only about 3 billion used for each word. It needs the memory of a 35B model and answers at close to the speed of a small one. On a Mac, that is often the best trade.
Which Mac mini to pick
| If you want to | Pick |
|---|---|
| Add a small local model to an agent that mostly uses the cloud | M4, 16 GB |
| Run useful mid-size models for tool calling and documents | M4, 24 or 32 GB |
| Run 30B-class models fast, next to other work | M4 Pro or M5 Pro, 48 GB |
| Run 70B models | M4 Pro or M5 Pro, 64 GB |
For local models, more memory beats a faster chip. A 32 GB M4 runs models that a 24 GB M4 Pro cannot load. When the memory is the same, the Pro chip and the newer generation answer faster.
Getting started
The simplest way is Ollama. It downloads models, runs them and gives your agent a local API. On Apple silicon it can use Apple's MLX framework, which is built for these chips.
brew install ollama
ollama serve &
ollama run qwen3.5:9bModel tags change, so check the exact name in the Ollama library before you pull.
Other good options:
- LM Studio, if you prefer an app with a window.
- mlx-lm, Apple's own tool, for the fastest results and for fine-tuning experiments.
Using a local model inside an agent
Local models rarely replace the largest cloud models. They work best next to them.
A common pattern:
- 1The local model reads every incoming request.It decides what kind of request it is, extracts the data and handles the routine ones itself.
- 2Only hard requests go to a cloud model.Reports, judgement calls, long reasoning.
Most requests in a real agent are routine, so most of them never leave the Mac. The cloud bill shrinks, and simple answers come back faster.
OpenClaw supports Ollama as a provider, and most agent frameworks can talk to its local API at http://127.0.0.1:11434.
Keep that port private. Ollama listens only on the Mac by default. Don't change it to listen on every network interface on a machine that is reachable from the internet. Reach it through an SSH tunnel instead.
- Chat with your accounting app on WhatsApp, the same split, in production
Why a hosted Mac mini for local models
- Always on. The model is loaded and ready at any hour, with no laptop to wake.
- Every memory size, without buying one. Try a 24 GB M4 for a day, move to a 64 GB M5 Pro when you outgrow it.
- Your data stays on your Mac. Requests a local model answers are never sent to an AI provider.
- Hosted in the EU.
Next