A trained language model is, in practice, a very large set of numbers called weights or parameters, plus a small amount of code that multiplies them with the numbers a prompt turns into. The weights are what training produces. Two other things sit upstream of them: the training data the model learned from, and the training code that ran the learning. A release can include any combination of the three, and “open weights” means only the first: the weight files can be downloaded and run on your own hardware. The Open Source Initiative’s definition of open source AI, version 1.0, asks for more. It names three components that together make up the preferred form for modifying a system: data information, meaning enough detail about the training data for a skilled person to build a substantially equivalent system; the complete source code used to train and run the system; and the parameters. It also lists four freedoms, to use, study, modify and share the system for any purpose. Measured against that yardstick a release of weights alone is not open source AI, and that gap is why the phrase “open weights” exists as a separate, narrower claim.
A download is not a permission, so the license comes next. Hugging Face, a public hub for model releases, has authors declare the license in the model card’s metadata, using an identifier such as apache-2.0, mit or llama4; for a license it has no identifier for, the author sets license: other, names it, and adds the license text to the repository. The two most common are short. The Apache License 2.0 grants every recipient a copyright license and a patent license, and asks them to pass on a copy of the license, mark files they changed and keep the notices; its patent grant ends for anyone who sues alleging that the work infringes a patent. The MIT license lets anyone use, copy, modify, publish, distribute, sublicense and sell the software free of charge, on the single condition that the copyright and permission notices stay with it. Neither text says anything about training data, and neither limits what the model may be used for. Qwen’s 235-billion-parameter model and OpenAI’s gpt-oss-120b both list Apache-2.0 on their model cards, and Mistral’s Mixtral 8x7B paper says its models were released under the same license.
Other labs write their own. The Llama 4 Community License Agreement is a custom license with conditions an Apache or MIT license does not have. A company whose products or services had more than 700 million monthly active users in the month before the Llama 4 release date must request a license from Meta, which may grant or refuse it at its sole discretion, and until then has no rights under the agreement. Every user must follow Meta’s Acceptable Use Policy, which sets out five categories of prohibited use, among them breaking the law, activities that risk death or bodily harm, and deceiving or misleading others. Anyone who distributes the materials or a product built on them must display a “Built with Llama” notice, and a distributed model created or improved using the materials or their outputs has to carry a name that begins with “Llama”. Hugging Face’s license list gives the Gemma Terms of Use and a Grok 2 community license identifiers of their own as well, so a custom license is common enough to have its own shelf. The word open, in other words, does not tell you which of these you are looking at. The license text does, and reading the primary source beats trusting a headline’s summary of it.
Access is a separate layer. A repository on Hugging Face can be gated: its authors turn on access requests, and a reader must be logged in and agree to share a username and email address with the authors, unless they opted out of collecting them. Approval can be automatic, so anyone who agrees gets in at once, or manual, so the authors decide. The authors can add their own fields, such as a company name, a country, or a checkbox confirming non-commercial use, and the documentation’s example even shows a flag for refusing visitors from the European Union where a license prohibits distribution there. The documentation also says the authors keep complete control and can block access at any time, even after approving a request. A gate records who asked and what they agreed to. It is not a license, and it says nothing about what the weights can be used for once they are on your disk.
Why would a lab publish weights at all, and why would it hold them back? Meta’s account of Llama 3.1 gives its reasons for publishing: developers can customize the models, fine-tune them on new datasets and run them on-premises, in the cloud or on a laptop without sharing data with Meta, and Meta argues that open models spread the benefits of AI more widely. The US National Telecommunications and Information Administration weighed both sides in a report dated Jul 30, 2024. Among the benefits it lists are a wider range of actors in AI research and development, including less-resourced ones, less market control by a few large developers, and the ability to work with a model without sharing data with a third party. Among the risks it lists are misuse that harms national security, safety, privacy or civil rights, weak oversight and unclear accountability. Its conclusion was that the evidence then available did not show restrictions to be warranted, nor that they would never be, and it recommended monitoring. The commercial side is simpler. The labs quoted in the lesson on how AI companies make money sell the use of a model by the token, and a customer who holds the weights can run the model without buying tokens. That is a plain reason to keep them private, though whether it is any given lab’s reason is for that lab to say.
So when a release is described as open, there are three questions that settle what you have. Can you download the weights, and under which license? Is there any description of the data and the code, or only the weights? And is the repository gated, with terms you are agreeing to when you press the button? A model that answers the first yes, the second no and the third no is open in the narrow sense that matters most to someone who wants to run it, and closed in the sense that matters to someone who wants to reproduce it. How a model is named and sized once you have it is the subject of the next lesson.