Why a small network (<2k parameters) exceed the limit of Keras optimizer?

Question

I\'m working on a very small model with less than 2K parameters:

Model: "model"
____________________________________________________________________