The approach I normally take is: I have small benchmarks for myself which test for things I care about.

And that has any models that I'm considering through that.