It has long been suggested that parametric polymorphism might be implemented by generating a distinct version of the code for each monomorphic instance of a function that is used in a given program. However, most implementations avoid this approach for fear that it could lead to a substantial increase in the size of the compiled program -- a so-called code explosion -- and that it would complicate the process of separate compilation. An alternative, used in many systems, is to have a single implementation of each polymorphic function and use a uniform (or boxed) representation for all values. This avoids the risk of code explosion but makes it more difficult to support parametric overloading. In addition, the use of boxed representations for even the most primitive values may carry a significant run-time cost. This paper presents a new approach that lies between these two extremes. The resulting system can be used to obtain an efficient implementation of overloading and to avoid the ...