Of course I'll cast my -1 vote as well.

I'll add to point 1:

If the data contained in the model is all of basic types or implicitly shared types then it's possible to already use the current framework without ever triggering any expensive* copy

, to point 2:

There's no clarity on how to support different types in different roles would we need an additional template argument for each used role beyond Qt::UserRole? is so the problems above grow exponentially with the roles

And to point 3:

No possible implementation for a reliable dataChanged emission if model internals are exposed directly

*expensive = a copy that implies more that a:

memcpy of less than 65bits and an increase/decrease of a numeric reference counter