Which AI model should I install on my Mac?

[Your name], founder of WorkOffline · Updated
Quick answer

Pick by memory. With 8 GB, use a small model of about 4 billion parameters (a download of about 2.5 GB). With 16 or 24 GB, an 8 billion parameter model (about 5.0 GB) is the best all-rounder. With 32 GB or more on an Apple Silicon Mac (M1 or later), a 14 billion parameter model (about 9.0 GB) gives better answers. Older Intel Macs can run the smaller models, but answers come more slowly.

You do not need to understand AI to answer this one. The single number that matters most on a Mac is memory (also called RAM or unified memory). The second is whether your Mac has an Apple Silicon chip (M1, M2, M3, M4 and later) or an older Intel processor. Get those two facts and the choice is quick.

Check your Mac in 30 seconds

  1. Click the Apple menu in the top left corner of the screen.
  2. Choose About This Mac.
  3. Read the Chip line (for example "Apple M2") or the Processor line (which says Intel on older Macs).
  4. Read the Memory line, for example "16 GB".

Memory is not the same as storage. A Mac with "512 GB" usually means storage (disk). Memory is the smaller number, typically 8, 16, 24, 32 GB or more.

The answer by memory

Model sizes are described in parameters, the number of settings the model learned during training. More parameters usually means better answers, but also a bigger download and more memory in use. The models below are open models from the Qwen3 family, compressed to about a quarter of their original size (a common format called 4-bit or Q4) so they fit on a laptop.

Recommended offline AI model size by Mac memory
Your Mac's memoryModel to installDownloadWhat to expect
8 GBAbout 4 billion parameters (Qwen3 4B Instruct 2507)About 2.5 GBQuick, good for drafting, summarising and questions about your files. Close heavy apps for best speed.
16 GBAbout 8 billion parameters (Qwen3 8B)About 5.0 GBThe sweet spot. Better writing and reasoning, still responsive.
24 GBAbout 8 billion parametersAbout 5.0 GBSame model as 16 GB, with lots of room left for your other apps and longer documents.
32 GB or more (Apple Silicon)About 14 billion parameters (Qwen3 14B)About 9.0 GBNoticeably better on tricky questions. Replies more slowly than the 8 billion model.
Any Intel MacAbout 4 billion, or 8 billion with 16 GB or more2.5 to 5.0 GBWorks, but slower. Short questions are fine; long summaries take patience.

Download sizes are the 4-bit (Q4_K_M) files published on Hugging Face, see Sources. Your Mac may show slightly different numbers because it counts gigabytes in a different way. For the full memory and disk picture, see how much memory and disk space offline AI needs.

Apple Silicon vs Intel: why it matters

On an Apple Silicon Mac, the processor and graphics chip share one pool of memory. AI models run much faster on the graphics part of the chip, and because the memory is shared, a Mac with 32 GB can hold a model that would need an expensive graphics card on a Windows PC. That is why Apple Silicon laptops are unusually good at running AI offline.

Intel Macs (most Macs sold before late 2020) have to run the model on the main processor. It still works, and the model fits the same way, but each answer takes longer. If your Intel Mac has 16 GB, you can try the 8 billion parameter model, but many people prefer the smaller one for speed.

Getting the best from it

  • Plug in for big jobs. Summarising a long report works the chip hard. On battery, expect it to drain faster while the answer is being written.
  • Close memory-hungry apps such as video editors, virtual machines or dozens of browser tabs, especially on 8 GB.
  • Leave disk space. Keep a few gigabytes free beyond the download itself. macOS needs room to work.
  • Expect a short pause at the start. The model loads into memory the first time you ask something. After that it is quicker.
  • Know the limits. Small models are good at writing and at working with your own documents, and weaker on general facts. See what a small AI model is.

How WorkOffline handles this

You do not have to pick by hand. On first run, WorkOffline reads your Mac's memory, chip and free disk space, recommends Lite, Standard or Large, and downloads that model once on Wi-Fi. After that it works with no connection. It tells you why it chose that size, and on an Intel Mac it says up front that answers will be slower. Choosing and swapping models from the full model library is a Pro feature. The details are in how WorkOffline picks a model.

Every ticket includes chat with your files, writing help and compress for email, priced at one flat white. See tickets and prices.

Frequently asked questions

Can an 8 GB MacBook Air run AI offline?

Yes. A model of about 4 billion parameters needs roughly about 4 GB of memory while it answers, which fits on an 8 GB Apple Silicon Mac. Close big apps such as video editors first and it will feel quicker.

Is 16 GB enough for offline AI on a Mac?

Yes. 16 GB comfortably runs an 8 billion parameter model, which is the best balance of quality and speed for most people. It uses roughly about 8 GB while answering.

Should I run the biggest model my Mac can fit?

Not always. A bigger model gives better answers but replies more slowly and leaves less memory for your other apps. If you want quick replies while you also have a browser and Office open, one size down is often nicer to use.

Does offline AI work on an Intel Mac?

It can, but the model runs on the processor rather than the graphics chip, so answers are noticeably slower. Stick to the smaller models on an Intel Mac.

Will running AI damage my Mac?

No. It works the processor hard, like exporting a video, so the fans may spin and the battery drains faster while it answers. It stops when the answer is done.

Sources

An evergreen guide from WorkOffline. No tracking scripts on this page.