Someone tried to modify the C source to Python to use the least significant bit of a PyObect* and have immediate integers. It didn't make much difference to performance.
Maybe it was Jeff Epler? He's nuts enough to have tried something like this :-)
Also, note that not every architecture in the world is byte addressed (though this is less of an issue if you use high bits as tags, but then you need more control over how your address space is layed out than you get from C).