The Model is quite simple:

var model = keras.Sequential(new List<Tensorflow.Keras.ILayer>()
{
    keras.layers.InputLayer(128),
    keras.layers.Dense(64),
    keras.layers.Dense(6),
    //keras.layers.Softmax(-1)
});

When using Softmax, it does not converge, while it converge well after removing it.

In tensorflow (python), the it's just the opposite. The model converge with softmax but does not without softmax.

Both in csharp and python I used sparse_categorical_crossentropy loss and Adam optimizer.