What is the difference between open-weight and closed AI models?

Every AI model is, at bottom, a very large file of numbers called weights, learned during training. The difference between open and closed is about who holds that file. With a closed model, such as the ones behind most big chat apps, the weights never leave the company. You send text over the internet, get an answer back, and your rights come from a terms-of-service agreement. With an open-weight model, the company publishes the file. You can download it, run it on your own computer or server, fine-tune it and, depending on the licence, ship it inside your own product.

In October 2026 both kinds keep improving, and many labs release some models of each. Choosing between them is covered in our earlier guide on open-weight versus hosted models. This article focuses on the part that trips people up: what the licence lets you do once you have the file.

Open weights are not the same as open source

"Open source" has a precise meaning in software, and the Open Source Initiative (OSI) extended it to AI with version 1.0 of its Open Source AI Definition in October 2024. To qualify, a system must let anyone use it for any purpose, study how it works, modify it and share it. To make those freedoms real, three things must be available: the parameters (the weights), the complete code used to prepare the data and train and run the model, and enough data information for a skilled person to build a substantially similar system.

Most models called "open" publish only the weights and some running code. They do not release the training code, and they describe their training data only in broad terms. That makes them open-weight: very useful, often free, but not open source in the OSI sense. The distinction matters less for running a model than for studying or rebuilding it.

The three families of licences

Permissive licences let you do almost anything. The MIT licence fits in a few lines: you may use, copy, change, merge, publish, distribute, sublicense and sell the work, provided you keep the copyright and permission notice, and it comes with no warranty. Apache 2.0 grants the same kind of broad, royalty-free rights and adds three things: a patent licence from contributors (which ends for anyone who sues over patents in the work), a duty to mark files you changed and pass on any NOTICE file, and no right to the maker's trademarks.

Custom "community" licences are written by the model maker. They are usually free for most people but come with conditions. Meta's Llama 3 licence is a well-known example. It allows use, copying, modification and redistribution without royalties, but it also asks you to:

• display "Built with Meta Llama 3" on products that use it,
• start the name of any model you fine-tune from it and distribute with "Llama 3",
• follow Meta's acceptable use policy, which forbids things like presenting outputs as human-written,
• request a separate licence if your products had more than 700 million monthly users,
• avoid using the model or its outputs to improve any other large language model.

Closed models have no licence to the weights at all, only terms for using the service.

A worked comparison

What do Apache, MIT and custom licences let you do?

With MIT or Apache 2.0, you can use the model commercially, change it, combine it with other work and redistribute it, including as part of a closed product, without asking permission or paying. Your duties are light: keep the licence text and notices, and under Apache 2.0 say which files you modified. Apache 2.0 also gives you an explicit patent licence, which many companies prefer, and it does not let you use the maker's brand as if it endorsed you.

With a custom licence, read before you build. The grant is often generous, but the conditions are where the surprises live: attribution wording, naming rules for fine-tunes, use policies that are part of the contract, size thresholds for very large companies, and limits on using outputs to train other models. A fine-tuned model normally inherits the licence of its base model, so check the original, not just the copy you downloaded.

A five-minute checklist before you commit to any open model: find the licence name on the model page; open the actual LICENSE file, since tags can be wrong; look for a linked use policy; check the base model's licence if this is a fine-tune; and note any rule about outputs if you plan to train other models with them. None of this replaces a lawyer for a serious product, but it answers the everyday question of what you are allowed to do.