the URL in the endpoint configuration should point to a single model. So, you should return a single zip file with just one trained model. The model will be loaded into the default project. The name of the model will be set to the name of the zip file. You cannot run Rasa NLU with a model server and have multiple models.
If you want to use a project with multiple models, you might want to store them locally or in a cloud storage.
Just a note: We are going to remove projects soon. For more information please read this post. Feel free to comment with your ideas.