Most of the issues pointed out by int_19h would be handled at compile time during static analysis and wouldn't change much to the generated bytecode. I'm talking about strong typing, immutability by default, less error-prone syntax, tail calls, evaluation order etc...

Even replacing 256bit ints with arbitrary precision "bigints" wouldn't add too much of a cost if it's a native type of the underlying VM (as it should be for such an application IMO). It might even reduce code size by removing overflow tests.