Run AI on your own computer with Ollama: what you need and which models to pick

ChatGPT and Claude run on someone else's computers. Every question you ask travels to their servers, and the best features cost a monthly fee. But there's another way: you can run an AI model on your own computer. It's free, your words never leave your machine, and it keeps working on a plane with no Wi-Fi.

The easiest way to do that is Ollama. This guide covers what your computer needs, how to get Ollama running in about five minutes, and, most importantly, which free models will actually run well on the computer you have.

What your computer needs

You don't need a gaming PC. Here's the minimum, and what makes life easier:

• A recent system. Windows 10 (version 22H2) or Windows 11, a Mac with macOS 14 Sonoma or newer, or Linux.
• Memory (RAM): 8 GB at the very least, 16 GB to be comfortable. This matters more than anything else, because the whole model has to fit in memory while it works.
• Disk space: about 10 GB free to start. Each model is a separate download, from under 1 GB for the smallest to 20 GB and more for the big ones.
• A graphics card is optional. Without one, Ollama uses your processor and still works, just more slowly. With an NVIDIA or AMD card, or a Mac with an Apple M-series chip, answers come several times faster.

One Mac note: Macs with Apple's M1, M2, M3 or M4 chips are excellent for this, because the chip and graphics share one pool of fast memory. Older Intel Macs work too, but slowly.

How much memory do you have?

• Windows: open Settings → System → About and look for "Installed RAM".
• Mac: click the Apple menu → About This Mac and look for "Memory".

The one rule that matters

A model needs roughly its download size plus 2 to 4 GB of free memory. And "8 GB of memory" doesn't mean 8 GB for the AI: Windows or macOS, your browser and your other apps are already using some of it. On an 8 GB laptop, expect about 3 to 5 GB to be free for a model.

If you pick a model that's too big, it won't break anything, but it will crawl: your computer starts shuffling memory to disk and everything slows down. The fix is simply to pick a smaller model.

Step 1: Install Ollama

1. Go to ollama.com/download.
2. Mac: download the app, open the file and drag Ollama into your Applications folder. Open it once.
3. Windows: download and run OllamaSetup.exe. You don't need admin rights, and it starts by itself when it's done.
4. Linux: paste this into a terminal:

curl -fsSL https://ollama.com/install.sh | sh

Ollama then runs quietly in the background, ready whenever you need it.

Step 2: Chat with your first model

The easy way: open the Ollama app, choose a model from the list, and type a message. The first time you use a model, Ollama downloads it, which takes a few minutes depending on your internet. After that it opens in seconds.

The terminal way: open Terminal (Mac) or Command Prompt (Windows) and type:

ollama run gemma3:4b

That downloads Google's Gemma 3 (3.3 GB) and starts a chat. Type your question and press Enter. Type /bye to leave. To try another model, just swap the name, for example ollama run llama3.2.

Step 3: A few handy commands

You only need four:

• ollama list shows the models you've downloaded and how big they are.
• ollama ps shows what's running right now. If it says 100% GPU, you're getting full speed. If it says CPU, the model runs on your processor and will be slower.
• ollama stop gemma3:4b frees the memory a model is using.
• ollama rm gemma3:4b deletes a model to get the disk space back.

Which models can your computer run?

Here's what runs well at each amount of memory. The number after each name is the download size. The command is what you'd type, or pick the same name in the app. All of these are free.

8 GB of memory: small but surprisingly useful

Great for writing help, summaries, questions and simple tasks. Expect answers at roughly reading speed without a graphics card.

• Gemma 3 4B (3.3 GB) gemma3:4b. Google's all-rounder. Good at writing, many languages, and it can look at pictures too.
• Llama 3.2 3B (2.0 GB) llama3.2. Meta's small model and one of the most downloaded. A solid everyday choice.
• Phi-4 mini (2.5 GB) phi4-mini. Microsoft's small model, quick and good at maths and step-by-step thinking.
• Qwen 3 4B (2.5 GB) qwen3:4b. Strong at reasoning for its size, and handles many languages well.
• Qwen 3.5 4B (3.4 GB) qwen3.5:4b. A newer Qwen that can also read images.
• DeepSeek-R1 1.5B (1.1 GB) deepseek-r1:1.5b. Shows its thinking before it answers. Fun to watch, and tiny.
• Gemma 3 1B (0.8 GB) gemma3:1b. For older or very low-memory computers.

16 GB of memory: the sweet spot

Noticeably smarter models that write better, follow instructions more closely and make fewer mistakes.

• Qwen 3 8B (5.2 GB) qwen3:8b. A favourite in this size, good at almost everything.
• DeepSeek-R1 8B (5.2 GB) deepseek-r1:8b. Thinks through problems step by step. Good for maths and logic.
• Qwen 3.5 9B (6.6 GB) qwen3.5:9b. Newer, reads images and can switch its thinking mode on and off.
• Gemma 4 12B (7.6 GB) gemma4:12b. Google's newest generation, good at reasoning, coding and pictures.
• Gemma 3 12B (8.1 GB) gemma3:12b. Well tested and dependable, with image understanding.
• Qwen 3 14B (9.3 GB) qwen3:14b. The most capable model you can comfortably fit here.

32 GB, or a graphics card with 24 GB: close to the paid assistants

This is where local AI starts to feel like the big services for everyday work.

• gpt-oss 20B (14 GB) gpt-oss:20b. OpenAI's own open model. Strong reasoning and good at using tools.
• Qwen 3.5 27B (17 GB) qwen3.5:27b. Very capable all-rounder that reads images too.
• Gemma 4 26B (19 GB) gemma4:26b and Gemma 4 31B (20 GB) gemma4:31b. Google's larger models.
• Qwen3-Coder 30B (19 GB) qwen3-coder:30b. Built for writing and fixing code.
• DeepSeek-R1 32B (20 GB) deepseek-r1:32b. Careful, step-by-step reasoning.

64 GB or more: workstations and big Macs

• DeepSeek-R1 70B (43 GB) deepseek-r1:70b.
• gpt-oss 120B (65 GB) gpt-oss:120b.
• Qwen 3.5 122B (81 GB) qwen3.5:122b.

Bigger than your computer? Borrow Ollama's

Ollama can also run the biggest models on its own servers, with the same commands. Add -cloud to the name, for example ollama run gpt-oss:120b-cloud. There's a free allowance to try it, and a Pro plan at $20 a month. Remember that cloud models send your words to Ollama's servers, so they're not private in the same way. Everything that runs on your own computer stays free and unlimited.

When it feels slow

• Try a smaller model first. A small model that fits beats a big one that doesn't. This is the most common mistake.
• Close browser tabs and heavy apps such as video calls. That frees memory for the model.
• Check ollama ps. If it shows mostly CPU on a computer with a graphics card, the model is too big for the card's memory. Go one size down.
• Run one model at a time. Use ollama stop before switching.
• Very long chats use more memory. Start a new chat when you change topic.

As a rough guide from people's tests, a mid-size model on a laptop's processor alone writes a few words a second. The same model on a graphics card, or a recent Mac, is five to ten times faster.

Is it really free?

Running models on your own computer costs nothing and has no limits. You don't need an account or a card. Each model does come with its own licence, though. Some, like Qwen, DeepSeek-R1 and gpt-oss, use standard open licences that allow business use; others, like Llama, come with their maker's own terms. If you plan to build a product on one, read the licence on its Ollama page first.

What next?

Ollama's app is deliberately simple. If you'd like a fuller chat window that looks and works like ChatGPT, with saved chats, document uploads and accounts for your family or team, add Open WebUI on top. And if you're curious how Ollama compares with LM Studio and llama.cpp, our Ollama page has a side-by-side comparison.