The short version: Ollama is a free program that lets you run AI on your own computer, privately. Think of it like a record player. Ollama is the player. The "models" are the records. You pick a record for the job you want done.
What is Ollama, in plain words?
Ollama is free software for Windows, Mac and Linux. You install it like any other program. Then you download an AI "model," which is the actual brain file. After that, everything runs on your computer.
- Private: your words never leave your PC.
- Free: no monthly bill.
- Works offline: no internet needed once a model is downloaded.
Ollama has its own simple chat window, so you can type questions right away. Step by step setup is in our Windows install guide.
What is a "model"?
A model is the file that holds what the AI learned. Different companies make different models, and each has its own strengths. Some are better at writing, some at pictures, some at code.
What do the numbers mean? (size vs speed vs smarts)
You will see names like llama3.2:3b or qwen3:14b. The "b" stands for billion. It is the number of "parameters," which is a rough measure of how much the model learned.
| Size | Speed | Smarts | Memory needed (roughly) |
|---|---|---|---|
| 1B to 4B (small) | Fast, even on older PCs | Fine for simple questions and short writing | Fits on 8GB computers |
| 7B to 9B (medium) | Good on most modern PCs | The sweet spot for everyday help | 16GB is comfortable |
| 12B to 14B (large) | Slower without a graphics card | Noticeably smarter and more careful | 32GB, or a 12GB+ graphics card |
| 27B to 32B (very large) | Slow unless you have a strong graphics card | Close to the online services for many tasks | 32GB+ and a 16GB to 24GB graphics card |
A simple rule: a model needs about 0.6GB of memory for every billion parameters in the usual download size. More detail in How much RAM and VRAM for Ollama.
Which model does what
These are popular families in the Ollama model library, checked October 2026. New ones arrive every few weeks. The code name is what you type to download it.
| Model family | Sizes | What it is good at | Speed | Smarts | Computer memory |
|---|---|---|---|---|---|
Llama 3.2 (Meta)llama3.2 | 1B, 3B | Small, friendly all rounder. Good first model on an older PC. | Fast | Basic | 8GB |
Llama 3.1 (Meta)llama3.1 | 8B | Solid everyday helper for writing and questions. | Medium | Good | 16GB |
Qwen 3 / Qwen 3.5 (Alibaba)qwen3, qwen3.5 | 0.6B to 235B | Very capable all rounders. Qwen 3.5 can also look at pictures. Many languages. | Varies | Good to very good | 8 to 32GB+ |
Gemma 3 / Gemma 4 (Google)gemma3, gemma4 | 1B to 31B | Clear writing and can read pictures. Gemma 4 is the newest. | Varies | Good to very good | 8 to 32GB |
Mistral (Mistral AI)mistral | 7B | Quick, plain spoken helper. Mistral Small 24B is the bigger sibling. | Fast | Good | 16GB |
Phi 4 / Phi 4 mini (Microsoft)phi4, phi4-mini | 3.8B, 14B | Small but good at math and reasoning. | Fast (mini) | Good | 8 to 16GB |
DeepSeek-R1deepseek-r1 | 1.5B to 671B | A 'thinking' model. It reasons step by step before answering. Slower, better for tricky problems. | Slow | Very good at logic | 16 to 32GB (8B to 14B) |
LLaVA, Qwen VL, Gemma 3 (vision)llava, qwen3-vl, gemma3 | 2B to 34B | Look at a photo and describe it, or read the text in it. | Medium | Good | 8 to 16GB |
Qwen 2.5 Coder, Qwen 3 Coderqwen2.5-coder | 0.5B to 32B | Writes and fixes computer code. | Medium | Good at code | 16 to 32GB |
Embedding modelsnomic-embed-text, embeddinggemma | Tiny | Not for chatting. They help apps search your own notes and documents. | Very fast | N/A | Any |
What fits on my computer?
- 8GB of memory: stick to small models:
llama3.2:3b,gemma3:4b,phi4-mini,qwen3:4b. - 16GB: the 7B to 9B models run well:
llama3.1:8b,qwen3:8b,mistral,deepseek-r1:8b. - 32GB: step up to 12B to 14B:
gemma3:12b,qwen3:14b,phi4,deepseek-r1:14b.
Not sure how much memory you have? On Windows, press Ctrl + Shift + Esc, click Performance, then Memory. The number at the top right is your total.
Pick by task
| I want to... | Try first | Also good |
|---|---|---|
| Chat, questions, writing emails | llama3.1:8b or qwen3:8b | llama3.2:3b on an 8GB PC |
| Read a letter or describe a photo | gemma3:4b or qwen2.5vl:7b | llava:7b |
| Math, planning, tricky logic | deepseek-r1:8b or phi4 | deepseek-r1:14b if you have 32GB |
| Help writing or fixing code | qwen2.5-coder:7b | qwen2.5-coder:14b with 32GB |
| Translate or other languages | qwen3:8b | gemma3:12b |
| Old or slow computer | llama3.2:3b or phi4-mini | gemma3:1b |
| Search your own documents | nomic-embed-text (plus an app like Open WebUI) | embeddinggemma |
Special helpers worth knowing about
- "Thinking" models like DeepSeek-R1 show their reasoning before the answer. Good for puzzles and planning. Slower.
- Vision models can look at a photo you give them. Handy for reading small print on a label or a letter.
- Embedding models work behind the scenes so an app can search your own files. You do not chat with them.
- One model per job is fine. Many people keep a small fast model for quick questions and a bigger one for careful work.
Your first 3 steps
- Download Ollama from ollama.com and install it.
- Open it and choose a small model like
gemma3:4borllama3.2:3b. - Type a question in plain words. That is it.
Optional, only if it helps
Gear that can make this easier
You do not need any of this to get started. These are simple add-ons that people often find helpful. Read reviews and check the return policy first.
Small mini PC for private AI
A quiet little computer that can run local AI like Ollama, if your current PC is too old.
See options on AmazonCommon questions
Is Ollama free?
Yes. Ollama is free software. The models in its library are free to download. You only need a computer to run them on.
Does Ollama send my chats to the internet?
When you use a model downloaded to your computer, the chat runs on your PC. Ollama also offers optional cloud models, which do go online, so pick a regular downloaded model if privacy matters.
Which Ollama model should a beginner start with?
On an 8GB computer, try llama3.2:3b or gemma3:4b. With 16GB, llama3.1:8b or qwen3:8b are good everyday helpers.
What does 8b or 14b mean in a model name?
It means 8 billion or 14 billion parameters, a rough measure of the model's size. Bigger models are usually smarter but slower and need more memory.
Who writes this: Blake, the guy behind AI Pick Vault. He has spent 47 years fixing things and sorting out PCs for regular people. He is 56 and still learning AI too, so these pages explain things the way he wishes someone had explained them to him.
Affiliate note: as an Amazon Associate I earn from qualifying purchases, at no extra cost to you. Prices change constantly, so we do not list them. Click through for the current price.