Definitely have to use each model for your use case personally, many models can train to perform better on these tests but that might not transfer to your use case.