It's definitely a cool idea. The only problem is that when it comes to double-bonds and triple-bonds, you will need different "atoms" fabricated with different orientation of flat-surfaces and magnets.
Take for example CARBON. The traditional 4-bond pyramidal orientation puts the 4 bonds as all going out in equally distributed directions in 3D around the atom. That's great for methane, ethane, butane... etc.. But say you want a double-bond on the carbons to make ethylene (2 bonds join the carbons), or acetylene (which has 3 bonds joining the carbons).
In those cases, you need to have different positions of the "flat" sides of the balls, to reflect the geometry of the bonds in space, and different numbers of magnets (and even different strengths if you want to simulate how much "energy" it takes to break and form the bond).
Then you have NITROGEN. It can form 3 bonds, but when bound to itself to form nitrogen gas, all 3 bonds are on the same side to link to the other nitrogen. This is very different from ammonia. And once again, 3-dimensionally you can't just stick 3 bonds at 120 -degrees apart around the equator. There is a slight bend towards once side due to electron cloud density on the top which warps it.
It's a great start, but really suited to very basic learning. You really need something in virtual reality here with gloves and 3D VR goggles like a chemical MineCraft or something that Microsoft Hololens can do.