In size limited categories like 4k and 64k intros as well as "challenges" like this, the size is the final size of the executable. The productions are typically delivered as a single executable with no other files, so everything is bundled into the executable. In 4k and 64k there is usually some kind of compression algorithm used to compress everything that's possible, including the actual executable code.
The "assets" are procedurally generated as much as possible. A hand-written software synthesizer and the control data ("sheet music") can fit in to less than one kilobyte, which is a lot less than producing a few minutes worth of music otherwise. Same goes for the textures, 3d models, etc.
If you download the 14k .zip file available on that web page, inside you will find a 128 byte .com file. In this case, there is no sound and the textures are generated procedurally.
Hopefully you see this, unfortunately your comment is dead.
Because I think your comment has value I will quote it here, please let me know if you'd like it removed.
"You can also have procedural music.
Viznut (a legendary Finnish democoder) did some experiments to see how short you could code something that still sounds at least interesting:
https://www.youtube.com/watch?v=GtQdIYUtAHg
I also wrote piece of JavaScript that can play these programs in case you want to try create your own: http://www.bemmu.com/music/index.html "