Correct Normals

Try rendering the normal vectors using OpenGL. Foreach normal, draw a single line from the vertex in question, P1 , to the point P2 defined by:

P1 = vertex;
P2 = vertex+normal*c;

.. where c is some appropriate constant, suitable for the scale of your model (maybe 1/100th of the whole model size)

Then you will visually see the normals in your renderings, and can judge whether they are actually normals to the surface or something else :)

Could you post a screenshot, including normals visualization?

If you want your model to be less "gouraud-shaded" and more "flat-shaded" (gives a more "square" look) - you should have three normals per "dice corner". In your images there are only one normal per "dice corner".

I don't know if this is the reason for the "dark areas" of your sides, but it it might be worth a try.

http://en.wikipedia.org/wiki/Gouraud_shading

Couldn't you convert the model data you are reading, into some other data structure?

From: one vertex, one normal, shared by many triangles per dice corner
To: soup of triangles or polygons, where each vertex has it's own normal (no shared normals).

This could be done by going through the model data polygon-by-polygon (or even triangle-by-triangle), and producing new polygonal data from that.

Of course simpler to say than do :)

IIRC, Blender3d allows you to specify soft / hard normals per-face. I don't know how this affects the exported model, but presumably hard normals will duplicate the relevant vertices.

Edit: Otherwise, maybe look into creating a normal map and use per-pixel lighting? I don't know if any free modelers can create normal maps, but maybe this is worth taking a look into. On the downside, this will increase hardware requirements (I don't think you can do per-pixel lighting with anything less than OpenGL 1.3).