I believe so, in this baseline implementation. In fact, that's also true of operations on doubles (non-Smi) in unoptimized code, which are stored as "heap numbers". Avoiding these allocations is one of the things an optimizing compiler can do (and does do, for doubles).

That said, for short-lived BigInts, the allocations may not be as expensive as you might think, since the garbage collector is generational and short-lived objects won't leave the nursery.