While demoscene stuff is seriously impressive most are using a shit load of DirectX to do what they do in such a small size. They are not bare metal demos.
(DirectX + 4kb) vs. (OpenGL + Python + Pyglet + 500LOC)
=
4kb vs. (Python + Pyglet + 500LOC)
Long are gone the days when you could manage the framebuffer yourself :) Specially if you want to run the demo in modern OSes.