The Parigot encoding has exponential-size normal forms: but don't panic!  With a decent graph-sharing implementation of lambda calculus, they take linear space in memory.