Well yes, you have to use one of the newer mnist variants these days if you want to get anything meaningful. A linear classifier gets something like 87% on the original one.