What's the difference between open and closed models?

A bakery can sell you a slice at the counter, or a whole cake to take home and cut however you like. Neither one hands you the recipe. Artificial intelligence (AI) models are shared in roughly those same ways. And the word "open," which you'll see on a lot of them, can mean the cake or the recipe, depending on who's saying it.

A closed model is one you use through a company's app or service. You type your question, the company's computers run the model, and the answer comes back to you. You never get the model itself. It's simple, and someone else looks after it. The company decides what it can do, what it costs, and when it changes. Many of the best-known chatbots work this way.

An open-weight model is one whose trained numbers, called its weights, are published for anyone to download. With a capable enough computer, you can run it yourself, even with no internet connection, and adjust it for your own work. The Open Source Initiative (OSI), the group that looks after the definition of open-source software, describes weights as the final settings of a trained network, the numbers that decide how it turns your words into an answer.¹

What open weights usually leave out is the recipe: the data the model learned from, and the code used to prepare that data and train it. Without those, you can use the model, but you can't fully check how it was made, or rebuild it yourself. The OSI points out that weights alone fall short of the transparency many researchers and regulators want.¹³

In 2024 the OSI published its Open Source AI Definition. To meet it, a system has to come with the freedom to use it, study it, change it, and share it. It also has to come with what you'd need to do those things: the weights, the full code for training and running it, and enough detail about the training data that a skilled person could build something very similar. By that bar, publishing the weights alone doesn't make a model open source.²³

The difference shows up in three places.

• Cost: running an open model yourself can be cheaper at large scale, but you pay for the hardware and the upkeep.
• Privacy: an open model on your own computer keeps your data with you, while a service you use sees what you send it.
• Control: with a closed model, the company chooses when it changes; with open weights, you can keep the version you downloaded.

Neither one wins everywhere. A service is easier to start with. Running your own takes skill, hardware, and care.³

So when you see a model described as open, ask one plain question: what exactly was shared? The weights? The code? The data, or a detailed description of it? Each answer changes what you can do with it. Next time a headline calls a model open, which would you look for first?