That's correct, sorry.
For projects where the model is hosted on the cloud (text and numbers) you would be able to run their sb3 file.
But for projects where the model is on the local computer (images and sounds), as you say, you can't immediately run their project.
For sounds projects, there is a Scratch block to train a new machine learning model, so as long as the project includes a "train new machine learning model block" then you could run their sb3 file, as it means you will end up with your own local machine learning model using the student's training data.
But for images projects where the model is hosted on the computer, that isn't currently an option, sorry.