The best local AI model for your computer

Which AI model your computer can run comes down to one number: its memory. This guide shows how to find yours, which open models run well with 8, 16, 32 GB or more, and what to expect from each, as of November 2026.

Updated November 3, 2026 · 6 min read

Want an answer for your exact computer? Try the model finder: pick your Mac’s chip or your PC’s graphics card and see the best model and how fast it runs.

Step 1: find out how much memory you have

• Mac: Apple menu → About This Mac. The Memory line is the number you need. On Macs with Apple chips (M1 and later), the AI uses this same memory.
• Windows PC: Settings → System → About, under Installed RAM. If you have a graphics card from NVIDIA or AMD, also open Task Manager → Performance → GPU and note Dedicated GPU memory. On a PC with a graphics card, that’s the number that matters most.

Step 2: the 60% rule

A model needs about its download size in memory, plus room for the conversation (longer chats and documents need more) and for everything else your computer is running. A good rule: keep the model’s download under 60% of your memory. That’s the same line PetekMind Hub uses for its Best fit pick.

Going over it doesn’t break anything, it just gets slow. On a Mac, the computer starts swapping to disk and answers come out word… by… word. On a Windows PC, a model that doesn’t fit on the graphics card spills into regular memory and runs several times slower. Without a graphics card, a PC runs the model on its processor alone, which is fine for small models and slow for big ones.

Step 3: the best model for your memory

These are good, current open models from different makers, all free to download through Ollama. Sizes are the download sizes.

With 4 GB or less, try Gemma 3 1B (0.8 GB). It’s instant and fine for simple questions, but don’t expect much more.

Picking by task

• Everyday writing, email and summaries: Qwen 3.5 in the biggest size that fits. It’s also good at using web search.
• Photos, screenshots and scanned documents: Gemma 3 4B, or Gemma 4 (E4B or 12B). They can look at an image and answer questions about it.
• Maths, logic and planning: Phi-4 mini on small computers. On bigger ones, DeepSeek R1 8B or GPT-OSS 20B, which “think” through the problem before they answer, so replies take a bit longer.
• Speed above all: Llama 3.2 3B or Gemma 3 1B.

How fast will it be?

Bigger models answer better but more slowly. Speed depends mostly on how fast your computer can move data through its memory, which is why Macs with Pro, Max or Ultra chips, and PCs with a recent graphics card, run the same model noticeably faster than a basic laptop. A simple test: if replies come out slower than you can read them, try the next size down.

Also check your free disk space. Each model is a file the size listed above, and you’ll probably want to keep two or three. PetekMind Hub won’t start a download that would leave less than 5 GB free.

How to try a model

With Ollama installed, open a terminal and run the name from the table, for example ollama run qwen3.5:9b. It downloads the model the first time and then lets you chat with it in the terminal. To delete one you don’t need, run ollama rm qwen3.5:9b.

The easy way: let PetekMind pick

PetekMind Hub runs on your Mac or Windows PC. It installs Ollama for you, reads your computer’s memory and free disk, and shows a short list of suggested models with the right one marked Best fit. One click downloads it. You can still add any other Ollama model by name.

Then the PetekMind app on Android and iPhone lets you use that model from your phone over your Wi-Fi, so a 27B model on the computer in the corner becomes everyone’s assistant. Read how to use local AI from your phone or how to share it with your family or office.

Open models improve every few months. We update this guide when better ones come out, and the hub’s suggestions change with it.