This tool expects large numbers of variant sites in order to achieve decent modeling with the Gaussian mixture model. It's difficult to put a hard number on mimimum requirements because it depends a lot on the quality of the data (clean, well-behaved data requires fewer sites because the clustering tends to be less noisy), but empirically we find that in humans, the procedure tends to work well enough with at least one whole genome or 30 exomes. Anything smaller than that scale is likely to run into difficulties, especially for the indel recalibration.