This model has only ever seen examples of hand-drawn digits for the numbers 0, 1, 2, 3, and 4. That is the only thing it knows how to do.
The 28×28 grid of 1s and 0s the model actually reads.
Every real weight inside the trained network. Note that the model is fixed once it has been trained. The only way for it to get smarter is to replace it with a new model.