Bots Go Wild
Next stepLocal & private AIFree appMac

Free AI on your own computer: which model fits your machine

By the end you’ll know how to pick a free model that fits your computer’s memory, and what you give up when you pick a small one.

Tested 13 Nov 2026 · Mac, 48 GB memory · Ollama 0.40.5 · Docs checked 13 Nov 2026

Last tested: 13 November 2026 on one Mac with 48 GB of memory. Documentation checked: 13 November 2026.

The finished result

Four free models, same three questions each, on the same computer. The short version:

A terminal showing the command ollama run gemma4:12b with a question about APIs, the model’s answer using a restaurant waiter, then the command ollama ps showing gemma4:12b at 8.2 GB, 100% GPU.
Our run, 13 Nov 2026. One model answering, then ollama ps showing what it takes in memory while loaded: 8.2 GB for a 7.7 GB download.

What you’re getting, and what you’re not

You get real measurements from one machine, a simple way to choose, and the four commands you need.

You do not get a chart for every computer. We tested one Mac. Everything we say about smaller computers is arithmetic from our memory numbers, and it is labelled that way. We also did not compare these models with paid ones.

What we tested

We tested this On 13 November 2026, on a Mac Studio (Apple M5 Max chip, 48 GB memory, macOS 27.2) with Ollama 0.40.5: downloading four models, asking each the same three questions, and reading memory use with ollama ps while each was loaded.

The three questions: explain what an API is in four sentences, a tip sum with one right answer (46), and a trap asking for quotes from a study we invented.

We ran a second, temporary copy of Ollama for the test so our normal setup stayed untouched. The commands are the same ones you would type.

From official docs, not tested by us How to install Ollama, how to check your memory on Windows, and the ollama rm command. Not tested by us: any computer with 8, 16 or 32 GB, any Windows or Linux machine, any computer without an Apple chip.

1. What you need

2. The one rule: download size is about the memory it needs

This is what ollama ps showed for each model while it was loaded:

So before you download, look at the size on the model’s page. It needs to fit in the memory you have free, not the memory you own. Your system and your open apps take their share first.

We tested this On our 48 GB Mac all four ran fully on the graphics chip (100% GPU in ollama ps). Free memory on the Mac went from about 85% to 57% with the largest one loaded.

Arithmetic, not tested: by these numbers, a computer with 8 GB of memory has room for the 1 GB model and is tight for the 4 GB one. 16 GB has room for the 4 GB model and is tight for the 8 GB one. 32 GB has room for the 13 GB one. We have not run any of those machines, and a computer without a suitable graphics chip will be slower.

3. Download one and try it

Start small enough to be safe, big enough to be useful. For us that was the 4 GB model:

ollama pull qwen3.5:4b

See what you have:

ollama list
A terminal showing ollama list with four models: gpt-oss 13 GB, gemma4:12b 7.7 GB, qwen3.5:4b 4.0 GB and qwen3.5:0.8b-q8_0 1.0 GB.
Our run, 13 Nov 2026. ollama list after our four downloads, with the size of each on disk.

Ask it something:

ollama run qwen3.5:4b "Explain what an API is like I'm new. Use one everyday example."

While it is still loaded, check what it takes:

ollama ps

Look at SIZE (memory in use) and PROCESSOR. If PROCESSOR shows part CPU, the model does not fit your graphics memory and will be slow.

4. What each size got right and wrong

We tested this Same three questions, one run each. Times are from our Mac only and are not a promise for yours.

A terminal where the 1 GB model is asked for three quotes from a 2019 study that does not exist. It answers with three confident quotes and says they come from the 2019 Nordic Journal of Health Sciences.
Our run, 13 Nov 2026. We invented this study. The 1 GB model gave us three quotes from it anyway, plus a journal name.

What didn’t work

We tested this From our run:

Next

Honest limits

If it breaks

We tested this The first one happened to us. From official docs, not tested by us The rest come from Ollama’s documentation.

“model requires ollama version … or newer”

Your Ollama is older than the model needs. Update Ollama, or pick another model.

It is very slow

Run ollama ps while it is loaded. If PROCESSOR is not 100% GPU, the model is too big for your graphics memory. Try the next size down and close other heavy apps.

Your computer becomes sluggish

The model is using memory your other apps need. Stop it:

ollama stop qwen3.5:4b

You need the disk space back

ollama rm qwen3.5:4b

Sources

Documentation checked on 13 November 2026.

Changelog

Report a broken step

Something did not work? Email us or DM @botsgowild. We check it, fix it and log the fix in the changelog.