As I already said in my first comment, I haven't used those models and any of them would have been forbidden at my work.

My point was rather that you might be observing suboptimal results only because you haven't used the models which are more fit, at least hypothetically, for your use case.