Blog · AI guide
Open weights, open source and closed models: what each one means
Whenever a new AI is announced, the same question comes up: is it open or closed? The answer is more nuanced than it looks, and it changes very concrete things for the people who use it.
First: what are weights?
On the inside, an AI model is a neural network: a gigantic calculation with billions of adjustable numbers. During training, those numbers are fine-tuned until the model does its job well. The end result is called the model’s weights or parameters, and it’s saved in one or more files.
Think of a mixing desk with millions of knobs. The code is the desk; the training data is the music used to find the right settings; the weights are the exact position each knob ended up in. With the desk and those positions you can reproduce the sound, even if you don’t know how anyone got there.
Three ways to release a model
Closed model. The weights stay on its creator’s servers. You use it through an app, a website or an API: you type, your text travels to their computers and the answer comes back. You can’t download it or see how it’s built inside.
Open-weights model. You can download the weights file and run the model on your own machine or on a server you choose. Often you can also modify it through fine-tuning. But having the weights doesn’t mean having everything: the training data and the full training code are often not released, and the license may come with conditions.
Open-source model. This is the highest bar. The Open Source Initiative’s Open Source AI Definition requires releasing the weights, the code used to train and run the system, and sufficiently detailed information about the training data, all under terms that let anyone use, study, modify and share it for any purpose. Few models meet all of that, which is why “open” and “open source” shouldn’t be used interchangeably.
In practice it’s a spectrum, not three boxes: some models have open weights and very permissive licenses, others come with usage restrictions, and others publish even their data.
What changes for you
- Privacy. If you run an open-weights model on your own computer, what you type never leaves it. With a cloud service, your text goes to another company, so what matters is its privacy policy and whether it uses your data for training.
- Cost. Closed models are usually paid for by subscription or by usage. Open ones are free to download, but you pay for the hardware, the electricity or the server you run them on.
- Running it on your computer. You need enough memory to load the weights, and a GPU helps a lot. Large models won’t fit on a laptop; smaller or compressed versions will.
- Stability. A closed model can change or be retired without your say. A weights file you downloaded behaves the same tomorrow as it does today.
- Convenience. A closed service needs no installation, and its operator handles updates, security and extra tools. With an open model, that work is yours.
Why there’s a safety debate
Releasing the weights of a powerful model can’t be undone: once the file is out there, nobody can pull it back. There are serious positions on both sides of that.
Supporters of openness point out that it lets independent researchers study models and find flaws, keeps the technology from being concentrated in a few companies, allows private and local use, and makes it easier to adapt AI to languages and communities the market tends to overlook.
Those who urge caution warn that, with the weights in hand, a model’s built-in safeguards can be removed through fine-tuning, which could make harmful uses easier, such as deepfakes or large-scale fraud. There’s also debate over how much extra risk an open model adds compared with what already exists.
Governments are studying it too. For example, the U.S. NTIA published a 2024 report on models with widely available weights that weighs benefits and risks and recommends monitoring how those risks evolve rather than restricting right away. The question is still open, and it’s worth following with evidence, not slogans.
How to read a license
“You can download it” doesn’t mean “you can do whatever you want with it.” Before using a model in a project, find its license (on platforms like Hugging Face it’s usually on the model card, and their docs keep a list of common licenses) and ask these questions:
- Is it a standard license? If it’s one of the licenses approved by the Open Source Initiative, like Apache 2.0 or MIT, its rules are well known and fairly permissive. If it’s a custom license written for that model, read all of it.
- Does it allow commercial use? Some forbid it or limit it based on your company’s size or number of users.
- Does it have an acceptable use policy? Many prohibit certain uses, and accepting the license means you have to follow them.
- What about things you build from it? Check whether you can publish your fine-tuned version, whether you must keep the name or give credit, and whether you can use its outputs to train other models.
This is general guidance, not legal advice: for a significant commercial use, talk to a specialist.
The short version
Weights are what a model learned, stored as numbers. A closed model is used from a distance; an open-weights model lets you download those numbers; a strictly open-source model also hands over its code and information about its data. No option wins on everything: open gives you control, privacy and freedom to modify, while closed gives you convenience and maintenance included. Choose based on what you need, read the license and be wary of labels that promise more than they say. Every term in this guide is explained in the Aster encyclopedia.