Open and closed models

When people argue about "open" and "closed" AI, they are usually arguing about one thing: whether you can get your hands on the model itself, or only on a chat box connected to it. That sounds like a technical detail. It turns out to decide where your words go, what you pay, whether it works without internet, and who gets to change the rules later.

To follow the argument you need one new word, weights. Everything else in this lesson hangs off it.

Weights: the part training leaves behind

How AI is trained walked through the stages a model goes through before you ever chat with it. What all that training actually produces is a very large collection of numbers called weights. Weights are the learned settings of the model: billions of tiny dials, each nudged a little at a time during training, until together they make good predictions about the next word.

A good way to picture it is a recipe card versus a kitchen. The training is the kitchen: the ingredients, the equipment, the months of trial and error. The weights are the finished recipe card that came out of it. Hand someone the card and they can cook the dish. They cannot see which ingredients were tried and thrown out, or rebuild the kitchen from it.

The weights are the model. When a company says it "released a model," the thing that matters is whether it released the weights. If it did, anyone with a capable enough computer can download that file and run the model. If it did not, the only way in is through a door the company controls.

One more piece: the weights alone do nothing until a program loads them and feeds them your text. That program is the part that reads your words as tokens and hands back the predicted reply. The chat apps you already use are a polished version of that program, sitting on top of weights you never see.

Closed models: you visit, you never take it home

A closed model is one whose weights stay private. You reach it only through the maker's app, its website, or an API, which is a way for other software to send it messages and get answers back. ChatGPT, Claude, and Gemini are where most people meet closed models. You type, your words travel to the company's computers, the model runs there, and the reply travels back.

That arrangement has real advantages. The biggest models need racks of specialized machines that no small business could buy, and the company keeps them updated, patched, and fast. You get the newest version without lifting a finger. It also has real limits. The company decides the price, the rules, which version you get, and when an old version goes away. And every message you send leaves your building.

Think of a closed model like a restaurant. The food can be excellent and you never wash a dish, but you eat what is on the menu, at their prices, during their hours, and the kitchen door stays shut.

Open weights: four families, as of October 2026

An open weight model is one whose weights the maker publishes for anyone to download. Most are posted on Hugging Face, a public site that hosts model files along with a model card, a page that describes the model and states its license. A license is the legal terms that say what you may and may not do with it.

Models come in families, a maker's line of related models in several sizes. Here are four well known open weight families, with the license terms taken from each maker's own license page or model card as of October 2026. Notice how different the terms are.

• Llama, from Meta. Llama 4 ships under the Llama 4 Community License, Meta's own terms. You may use and build on it, but you must follow Meta's acceptable use policy, show "Built with Llama" on products that use it, and any company with more than 700 million monthly users has to ask Meta for a separate license.
• Gemma, from Google DeepMind. Gemma 4 is released under Apache 2.0, a standard permissive license that allows commercial use and changes with few strings attached. Earlier Gemma models came under Google's own Gemma Terms of Use, which include a prohibited use policy. Same family, different terms depending on the generation.
• Mistral, from Mistral AI. Mistral Large 3 is under Apache 2.0. Mistral Medium 3.5 uses a modified MIT license that allows commercial use but carves out an exception for companies with large revenue. One maker, two very different sets of terms.
• gpt-oss, from OpenAI. OpenAI, the company behind ChatGPT, publishes gpt-oss as open weight models under Apache 2.0. The same company sells closed models and gives away open ones, which tells you this is not a fight between two camps. It is a choice makers make model by model.

Those terms change between versions, so treat the list as a snapshot. The habit worth keeping is the one in the Try this yourself section: read the license line on the model card, every time.

Open weights also mean people can keep training a model on their own examples, which is the fine-tuning lever from the last lesson. That is why one popular open model often has hundreds of tweaked versions floating around, each made by someone else.

Running a model on your own computer

If you can download the weights, you can run the model on a machine you own. People call this running a model locally, meaning on your own hardware instead of someone else's servers. This lesson does not walk through setup. What matters here is the trade you would be making.

What you gain. Your words never leave the computer, so there is no company storing them, reviewing them, or training on them. There is no charge per message. It keeps working with the internet down. And the model does not change under you, because a file on your drive only changes when you replace it.

What you give up. Size, mainly. A model has to fit in your computer's memory to run well, and the biggest models are far too large for a home or office machine. The open models that fit on a laptop are the smaller members of each family, and smaller usually means weaker on hard reasoning, long documents, and niche facts. Replies can also come out noticeably slower than a chat app on data center hardware. And you are the one responsible for keeping it updated.

Tap each priority below to see how the two sides compare on the things people usually care about.

Here is how that plays out for a real decision. Picture a side by side rental shop in Hurricane that wants AI help summarizing customer damage notes and phone numbers. The owner's first question is not which model is smartest. It is whether those notes may leave the building at all. If the answer is no, a closed chat app is off the table no matter how good it is, and a smaller model running on the shop's own computer becomes the only honest option, even if its summaries are a little plainer.

That is the privacy argument for open models in one sentence: if the model runs where the data already is, the data never has to travel. Privacy for your files covers what happens to an upload in a closed app and which switches control it. Running locally sidesteps those questions entirely. The catch from the explainer applies, though: an open model you use through someone else's website is back on someone else's server, with their policy. Open weights only protect privacy when you run them yourself.

How much quality you give up is the part to test rather than assume. The final lesson, Evals and benchmarks, shows how to compare two models on your own questions.

Open weights is not the same as open source

In software, open source has a specific meaning: the code is published under a license that lets anyone use, study, change, and share it. When AI makers call a model "open," many people assume it means the same thing. Usually it does not.

The Open Source Initiative, the group that maintains the standard definition of open source, published the Open Source AI Definition, version 1.0. It says an AI system counts as open source only if three things are available under open terms: detailed information about the data used to train it, the complete code used to train and run it, and the parameters, which is the formal word for the weights. The bar is that a skilled person could build a substantially equivalent system from what was shared.

Weights are one of three. A release that publishes the weights alone, under any license, does not meet that definition. And some open weight licenses fall short in a second way: terms like a user cap, a prohibited use list, or a revenue carve out mean the license itself is not fully open, even before you ask about the data.

Why should anyone outside the tech world care? Because what a model learned from shapes what it says, which biases it carries, and which gaps it has. With a closed model you cannot look. With most open weight models you still cannot look. You simply get to run it where you choose.

Try this yourself

No downloads, no accounts. You are going to see how well a chatbot explains a model license. Pick one of the four families above and fill in its name, maker, and license from the list. If you want to check the license line yourself, it is shown near the top of each model's card on Hugging Face, but that step is optional. Then paste this into the chat app you already use:

I am looking at an AI model called [model name] from [maker]. Its model card says the license is [license name].

In plain English:
1. Are the weights downloadable?
2. Can a small business use it commercially?
3. Does this release count as open source under the Open Source Initiative's Open Source AI Definition? What would be missing?
4. If I ran it on my own computer, where would my prompts go?

Watch for two things. First, does the chatbot separate "free to use" from "open to inspect," or does it treat open weights and open source as the same thing? Second, compare its answer about commercial use with the license page itself. If they disagree, the license page wins, every time. Then ask the same chatbot whether its own weights are published, and notice that you are talking to a model you can only visit.