Except that Llama 4's license restricts you from using its multimodal features in the European union, and the Llama 3 license restricted using the model output to augment the training of other models (distillation). Both are important use cases. It very much isn't an academic exercise. It has a real impact on what you can (legally) do with a model.

With a model licensed using MIT, you know what you are getting. You know you can modify it. You know you can use it for any purpose whatsoever, including ones the corporation that built it dislikes. Using the MIT license is "no ace up my sleeve" licensing - there are no gotchas that will come and bite you later.