How much less energy does a query against an 8B model take, vs a 70B model? Can we get more clever about using smaller, more specialized models?