Published on August 7, 2025 1:04 AM GMT
Disagreements degenerate into debates about word usage way too often and such debates are usually pointless. However, sometimes words can mislead or provoke emotional reactions and in those cases word usage becomes important. This is one of those cases.
Many modern AI models, such as those typically used for text or image generation, consist mainly of a huge neural network. The AI models, for which parameters (a.k.a. weights) are freely distributed on the internet came to be known as open weight models. This is contrast to other models, whose parameters are kept secret on some servers and are only accessible through APIs or web interfaces.
Some[1] have instead used the term open source to designate the same thing. I cannot speak about the motivation behind that word choice, but it ends up distorting what "open source" means, creates confusion and could make publishing weights seem more virtuous than it is.
What is open source?
I won't get into all the details of defining open source. Here are a couple of points from the definition by the Open Source Initiative (the organization who introduced the term in 1998): [2]
2. Source Code
The program must include source code, and must allow distribution in source code as well as compiled form. Where some form of a product is not distributed with source code, there must be a well-publicized means of obtaining the source code for no more than a reasonable reproduction cost, preferably downloading via the Internet without charge. The source code must be the preferred form in which a programmer would modify the program. Deliberately obfuscated source code is not allowed. Intermediate forms such as the output of a preprocessor or translator are not allowed.
What is the source code of a generative model?
Let's say I want to use the model for my company, while making sure it knows nothing about competitors. The most reliable way to do that would be to remove such references from the training data and retrain the model. This illustrates that the preferred form for modifying a generative model, such as Llama or DeepSeek would include:
- source code used for training, andtraining data.
This is not what is provided when a model's weights are published so we should not call it "open source". There are a number of reasons why providing the training data is unrealistic and probably never going to happen. Retraining the model would cost a fortune anyway so one could argue that open source is just an unsuitable distribution method for generative models.
5. No Discrimination Against Persons or Groups
The license must not discriminate against any person or group of persons.6. No Discrimination Against Fields of Endeavor
The license must not restrict anyone from making use of the program in a specific field of endeavor. For example, it may not restrict the program from being used in a business, or from being used for genetic research.
Open weight generative models typically restrict usage based on company size and use case (e.g. Llama's license disallows use for military purposes and use by companies with over a given number of users).
It is unclear whether company size restriction breaks rule 5, but I would argue the restriction is against the spirit of the rule. The use case restriction certainly breaks rule 6.
Conclusion
While "open source" was defined before open weight generative models were a thing, the intuition behind the term points to something quite different from those models. Open weights very closely matches what was traditionally considered closed source - training corresponds to the compilation and the output is an inscrutable array of numbers, which can be used to perform some computation.
Publishing weights is often misrepresented as if it supports some open source ideology. In my opinion companies use it to signal things like "we're the good guys", "we care about our users".
The term "open source" is useful vocabulary and becomes meaningless if we use it when we actually mean "closed source".
Discuss
