dumb question - why not create smaller language-specific models?