A free program that runs an AI model on your own computer - no internet, no stranger's server reading what you type.
Ollama is a small command-line program (the text window where you type commands instead of clicking with a mouse) that downloads, runs and manages open-source AI models. Open source here means the model itself - the file that holds what it learned - is publicly available to download, unlike models such as OpenAI's GPT, which you can only reach through their site.
In practice it looks like this: you type one command, Ollama downloads a model like Llama (built by Meta, the company behind Facebook and Instagram) or DeepSeek (a Chinese AI company), it weighs a few gigabytes, and within minutes it's answering you in the same window. No browser, no account, no chat history stored by someone else. Your computer does the job a server on another continent would otherwise do.
It was created in 2023 by Jeffrey Morgan and Michael Chiang, two developers from Palo Alto, California, right when every major AI model required an account, an internet connection and sometimes a payment card. The idea is simple: open-source models already existed for free, what was missing was an easy way for an ordinary person to run them on their own machine without being an engineer.
Today it's used by developers testing AI features without a cloud bill, companies with sensitive data who don't want text leaving the office, and people who simply want AI to work without internet too. The company recently raised $65 million in funding, showing that local AI is now a serious market, not a hobby.
Here's the core of it
The difference is like between a home oven and a restaurant. At the restaurant you order, you wait, you pay every time, and you never see exactly what happens in the kitchen. With a home oven you buy the ingredients once and then cook whatever you want, whenever you want, no bill at the table and no stranger's eyes in your kitchen.
As someone who builds AI systems myself, that's exactly why I care about tools like Ollama. Not because cloud models are bad, but because control should be a choice, not a default. When the model is on your disk, you decide what goes in and out of it - not some company on another continent.