While traditional methods for calling variants across whole genome sequence data rely on alignment to an appropriate reference sequence, alternative techniques are needed when a suitable reference does not exist. We present a novel alignment and assembly free variant calling method based on information theoretic principles designed to detect variants have strong statistical evidence for their ability to segregate samples in a given dataset. Our method uses the context surrounding a particular nucleotide to define variants. Given a set of reads, we model the probability of observing a given nucleotide conditioned on the surrounding prefix and suffixes of length k as a multinomial distribution. We then estimate which of these contexts are stable intra-sample and varying inter-sample using a statistic based on the Kullback-Leibler divergence. The utility of the variant calling method was evaluated through analysis of a pair of bacterial datasets and a mouse dataset. We found that our variants are highly informa- tive for supervised learning tasks with performance similar to standard reference based calls and another reference free method (DiscoSNP++). Comparisons against reference based calls showed our method was able to capture very similar population structure on the bacterial dataset. The algorithm's focus on discriminatory variants makes it suitable for many common analysis tasks for organisms that are too diverse to be mapped back to a single reference sequence.