This longitudinal study assessed the relation between maternal mental state talk over time and the development of social understanding in young Pacific Island children growing up in New Zealand. At all three time points (15, 20, and 26 months), mother and child pairs (N = 45) engaged in a free play task and maternal talk was coded for mental state and non-mental state content. Mothers reported on children's language ability using the MacArthur Communicative Development Inventory - MacArthur-Bates Short Form Vocabulary Checklist (MCDI-SF) and a Mental State Language Form (MSLF). Children's social understanding was measured using two emotion understanding tasks (emotion recognition and emotion situation), two perspective-taking tasks (move object and occluded object), and a deception task (penny-hiding). Mothers' strength of ethnic identity was measured using an abridged version of the General Ethnicity Questionnaire (GEQ-Short). Results indicated that mothers' and children's mental state language increased over time. Mothers who identified more strongly as Pacific Island tended to use less mental state language than mothers who identified more strongly as Pakeha or NZ European, and had children with lower scores in social understanding tasks. In particular, mothers' use of emotion state language and self-and-other directed desire talk at 20 months was uniquely predictive of children's social understanding at 26 months. Also, children with 2 or more siblings had significantly better social understanding at 26 months. Working within a social constructivist framework and drawing upon Vygotsky's principle of 'zone of proximal development', I argue that primary caregiver's ethnic identity, their use of specific mental state language, and the presence of older siblings play a crucial role in children's development of social understanding in the first few years of life.