Guides · Generative AI

Run an LLM Locally with Ollama on Windows and Mac (2026): A Beginner's Guide

By Ravindra Bagale, trainer · LinkedIn · Published: 19 October 2026 · Last updated: 19 October 2026 · 8 min read

To run an AI model on your own computer for free, install Ollama from ollama.com/download (Windows 10 22H2 or newer, or macOS 14 Sonoma or newer), open a terminal and type ollama run llama3.2. It downloads a 2.0 GB model once and then chats with no internet and no account. On 8 GB of RAM, stick to small models of 1 to 4 billion parameters; they are slower and less clever than ChatGPT, but your text never leaves your computer.

1. Suppose Imran cannot paste office notes into ChatGPT

Suppose Imran works at a small accounts firm in Hyderabad. His office does not allow client notes in online chatbots. He still wants AI to tidy up meeting notes and explain error messages. His laptop has 8 GB of RAM and Windows 11. A friend tells him: "Run the AI on your own laptop with Ollama, and nothing goes online."

2. What you need (checked on 19 October 2026)

Models are separate downloads. On ollama.com/library, llama3.2:1b is 1.3 GB, llama3.2 (3B) is 2.0 GB, gemma3:1b is 806 MB and gemma3 (4B) is 3.4 GB. A common rule: about 8 GB of RAM for a 7B model and 16 GB for 13B.

3. Worked example 1: install and run your first model

On Windows:

1. Download OllamaSetup.exe from ollama.com/download and run it. No admin password is needed.
2. Open PowerShell or Command Prompt.
3. Type:

ollama run llama3.2

On Mac: open the downloaded app, drag Ollama into Applications, start it once, then open Terminal and type the same command.

The first time, it downloads the model (2.0 GB). Then you see a >>> prompt. Type:

>>> Explain in 3 lines what a balance sheet is, for a beginner.

Expected output (yours will be similar, not the same):

A balance sheet is a summary of what a business owns and owes on one date.
On one side are assets, like cash, stock and machines.
On the other are liabilities and the owner's capital, and both sides must be equal.

Type /bye to leave the chat. Now switch off the Wi-Fi and run ollama run llama3.2 again: it still answers.

4. Worked example 2: the 5 commands you will use

ollama run llama3.2:1b     download (first time) and chat with a smaller model
ollama ls                  list the models you have downloaded
ollama ps                  show models loaded in memory right now
ollama rm llama3.2:1b      delete a model to free disk space
/bye                       leave a chat

Expected output of ollama ls after both downloads (sizes from ollama.com; IDs and dates will differ):

NAME              ID      SIZE      MODIFIED
llama3.2:1b       ...     1.3 GB    2 minutes ago
llama3.2:latest   ...     2.0 GB    10 minutes ago

5. Worked example 3: call the model from a script

Ollama also runs a small local server at localhost:11434, so your own programs can use the model. Type this in Command Prompt on Windows (not PowerShell, which reads the quotes differently) or in Terminal on Mac:

curl http://localhost:11434/api/generate -d "{\"model\": \"llama3.2\", \"prompt\": \"Give 2 tips to write clear meeting notes.\", \"stream\": false}"

Expected output (shortened; the words will differ):

{"model":"llama3.2","created_at":"...","response":"1. Write the decision and the owner for each point. 2. Send the notes the same day.","done":true, ...}

Tools such as n8n and Python libraries can use the same address.

6. Worked example 4: pick a model for your RAM (Python)

Which model should you try first? This script picks the largest model from a short list that fits in about half your RAM, leaving the rest for Windows or macOS and your apps.

# Which Ollama model should I try first? Picks by your computer's RAM (model sizes from ollama.com/library, 19 Oct 2026).
MODELS = [  # name, download size in GB
    ("gemma3:1b", 0.8), ("llama3.2:1b", 1.3), ("llama3.2", 2.0), ("qwen3:4b", 2.5),
    ("gemma3", 3.4), ("qwen3:8b", 5.2), ("gemma3:12b", 8.2),
]
def pick(ram_gb):
    budget = ram_gb * 0.5          # leave about half the RAM for Windows/macOS and your apps
    fits = [m for m in MODELS if m[1] <= budget]
    return fits[-1] if fits else None

for ram in (4, 8, 16, 32):
    m = pick(ram)
    print(f"{ram:>2} GB RAM -> " + (f"ollama run {m[0]}  ({m[1]} GB download)" if m else "too little RAM; use a web chatbot"))

Real output (19 October 2026):

 4 GB RAM -> ollama run llama3.2  (2.0 GB download)
 8 GB RAM -> ollama run gemma3  (3.4 GB download)
16 GB RAM -> ollama run qwen3:8b  (5.2 GB download)
32 GB RAM -> ollama run gemma3:12b  (8.2 GB download)

Imran has 8 GB, so he starts with gemma3 (4B, 3.4 GB). If his laptop feels slow, he drops to llama3.2.

7. Local vs online AI: when to use which

• Use local (Ollama) for private notes, offline work, learning how models work and building small tools.
• Use online (ChatGPT, Gemini, Claude) for the best quality, web search, images and long documents.
• Local is not automatically safe: download Ollama only from ollama.com, and models only from its library or sources you trust.
• Check answers: small models make more mistakes than big online ones.

8. Beginner mistakes and quick fixes

9. Try at home

Recap

An LLM is a file and a program. With Ollama, both sit on your laptop, and your words stay there too.

Understand the model: What Is an LLM? Large Language Models Explained Simply · Next idea: n8n AI Agent Tutorial for Beginners

Ravindra Bagale, trainer: linkedin.com/in/ravindra-bagale. Imran and his office are made up for learning, and the chat and API answers are examples; a real model words them differently. System needs and commands were checked on docs.ollama.com, and model sizes on ollama.com/library, on 19 October 2026. The Python output was produced on 19 October 2026.

Related: Local AI in VS Code (Ollama + Continue)