How does this compare to just training separate quantized models? Is it actually easier in practice?