That point must depend on the RAM available on your computer and graphics card.
From my experience with WebGL which must work similarly to Java 3D / OpenGL, you load vertice, normals and texture coordinates buffers of each group in the graphics card, then request to draw the buffers with a transformation matrix which will place it at the spot.
Therefore, it's better to merge faces into a larger group. It will create a larger buffer in the graphics card and in the main RAM, but it will require less instructions to draw all the groups.