Every classic license comparison is framed by two questions: does the license force you to release your own work under the same terms (copyleft), and does it grant an explicit patent license? Run those two questions over software and you get a tidy map. Run them over the 200 open models in our dataset and the map rearranges itself: Apache 2.0 covers 87 models, MIT covers 29, and GPL, the license with the teeth in every software comparison, covers none at all. This page gives each classic license its section, puts them side by side with the model counts attached, and then answers the practical question the counts raise.
MIT License: maximum simplicity, 29 models
MIT asks for one thing: keep the copyright and permission notice. No copyleft, no patent clause, no warranty. On models it is the second most common classic license in the dataset, 29 models, plus 2 more under a Modified MIT whose modification is the interesting part: Kimi K2's version attaches a naming and display duty once a product passes 100 million monthly users, a reminder that "Modified" in a license name is where conditions move in. For word-for-word MIT, the notice is the whole job.
Apache License 2.0: permissive plus patent protection, 87 models
Apache 2.0 allows the same commercial use, redistribution and modification, and adds what MIT lacks: an explicit patent grant from contributors, with a retaliation clause that ends the license for a licensee who sues over the covered work, and a duty to state changes in modified files. It is the dataset's runaway default, 87 of 200 models, nearly three times MIT's count. The enterprise logic is the one every comparison cites, and the model world has ratified it: when a foundation or a lab releases weights it wants adopted everywhere, Apache is the license it reaches for. Our Apache vs MIT pair comparison isolates just these two; the counted fact is that the pair together covers 118 of the 200 models.
GPL: copyleft, with teeth, and zero models
GPL's bargain is the opposite of permissive: distribute a derivative work and the derivative ships under GPL, source included; GPLv3 adds a patent grant of its own. In software that bargain built an ecosystem. On model weights it has essentially not been taken up: zero of the 200 models in our dataset carry a GPL license, and zero carry BSD either. The likely reasons are structural. What counts as a derivative of a set of weights, a fine-tune, a merge, a distillation, is unsettled in a way that derivative code is not, and labs releasing models want adoption without that argument. Whatever the reason, the empty seat is a fact of the current landscape, and it was filled by something else: the custom community licenses (Llama on 23 models, Gemma Terms on 14) that borrow copyleft's instinct, control, without its text.
Side by side, with the models attached
| License | Copyleft | Patent grant | Attribution required | Models in our dataset |
|---|---|---|---|---|
| MIT | No | No | Yes (notice) | 29 (plus 2 Modified MIT) |
| Apache 2.0 | No | Yes | Yes (notice, plus state changes) | 87 |
| BSD (2/3-clause) | No | No | Yes (notice) | 0 |
| GPL (v2/v3) | Yes | v3 only | Yes | 0 |
How to use this in practice
Three buckets cover the classics: permissive without a patent grant (MIT, BSD), permissive with one (Apache 2.0), and copyleft (GPL). For model work the buckets decide less than the counts suggest, because the licensing risk on a real model rarely sits in its classic license at all: 192 of the 200 models allow commercial use, attribution is required on 198, and the conditions with consequences, user ceilings, prohibited-use policies, no-competing-training clauses, live in the custom licenses counted in our license types census and the restrictions census. Choose between MIT and Apache on the patent question if you are releasing; if you are consuming, read the model's own terms first and its classic license second.
Frequently asked questions
Can I mix MIT and Apache 2.0 models in the same product? In the dataset's terms, both allow commercial use, redistribution and modification, and neither imposes copyleft, so they coexist; Apache's notice and patent terms travel with its models, and each model's own attribution duty still applies.
Is GPL always incompatible with a commercial model product? The question rarely arises in practice, because no model in our 200-model dataset is GPL-licensed at all. The compatibility debates in model licensing happen around custom community licenses instead.
What is the real difference between BSD and MIT? In software, very little: BSD's extra clause mainly bars using the authors' names to endorse derivatives. On models the difference is simpler still: MIT appears on 29 models in the dataset and BSD on none.
Why does the Apache license matter so much for models? Because of the patent grant and the adoption signal: 87 of 200 models, including most foundation releases meant for commercial building, ship under it. When in doubt between two comparable models, the Apache one carries the fewer open questions.