Abstract. Next generation sequencing (NGS) technologies are being applied to many fields of biology, notably to survey the polymorphism across individuals of a species. However, while single nucleotide polymor- phisms (SNPs) are almost routinely identified in model organisms, the detection of SNPs in non model species remains very challenging due to the fact that almost all methods rely on the use of a reference genome. We address here the problem of identifying SNPs without a reference genome. For this, we propose an approach which compares two sets of raw reads. We show that a SNP corresponds to a recognisable pattern in the de Bruijn graph built from the reads, and we propose algorithms to identify these patterns, that we call mouths. We outline the potential of our method on real data. The method is tailored to short reads (typ- ically Illumina), and works well even when the coverage is low where it reports few but highly confident SNPs. Our program, called kisSnp, can be downloaded here: http://alcovna.genouest.org/kissnp/.