James Robertson asks the 64-bit question for Smalltalk.
Implementing dynamic languages with 64-bit words always had an appeal to me for this reason:
You can encode 64-bit floats as literals. Everything else can be encoded within a NaN. The result is that arrays of large floats are more efficient in space and time, enabling more efficient analytical computing. You can also encode more information in an immediate object, like characters with styles, full 32-bit integers without shifting tag bits, etc.