What's the difference between an open-source and a closed-source AI model?

The core difference is about access and control, not just price. A closed-source model is like a restaurant where you can only order from the menu. You get the meal, but you never see the recipe, the kitchen, or the ingredients. ChatGPT, Google's Gemini, and Midjourney are all closed-source. You use them through a website or an app, but you can't download the model, look at its code, or run it on your own computer. An open-source model is more like a cookbook that's been published for everyone. The actual 'weights'—the numerical values the model learned during training—are publicly available for download. Meta's Llama models and Mistral are famous examples. You can download them, run them on your own hardware, and even fine-tune them for a specific task, like making a customer service bot that only knows about your company's products. Here's where it gets a little blurry, though. 'Open-source' in the AI world is a bit of a messy term. Some companies release the model weights but not the training data or the exact code used to train it. That's like getting the cookbook but not knowing where the chef sourced the ingredients. For a beginner, the practical difference is this: closed-source models are much easier to start with—just open a browser. Open-source models give you total privacy, since your data never leaves your machine, and they're often cheaper to use at a large scale, but you'll need some technical skill to set them up. A good tip: if you're just experimenting, don't get hung up on this debate. Start with a free tier of a closed-source tool. If you later find you need more privacy or want to build a custom tool without paying per-query fees, that's the moment to look into running an open-source model on your own computer.