Strictly speaking, a model is made by a set of parameters which define the function that best fits the data with which the model has been trained. Each model has a given number of hidden units, and you can expect a model with a different number of hidden units to give different results. Models can also differ by the initial weights which have been used as guesses (this is controled via the parameter called seed).
But such individual models often have, like human being, domains in which they are particularly good, and others in which they are not. For this reasons what we call a model is actually a set of individual models with given numbers of hidden units and different seed. Their predictions are combined to give the best overall results possible.
When you use the model manager, you actually don't get to handle individual submodels, or seldom. Instead, you access your model via a name that you choose initially, for example "Creep_of_autenitics".