Split the model based on language family

[jesperpedersen]

Could it be an idea to split the model into smaller pieces ?

I mean if I use C (or C++) then I don't need a model that understand Python.

I think could help lower the xB requirements, and allow more people to load it locally