This article presents a new model of transcoding numbers from verbal to arabic form. This model, called ADAPT, is developmental, asemantic, and procedural. The authors' main proposal is that the transcoding process shifts from an algorithmic strategy to the direct retrieval from memory of digital forms. Thus, the model is evolutive, adaptive, and particularly suited to account for learning and development. ADAPT, which has been implemented in a computer program, accounts for existing and new developmental data, for new data in adolescents with mild retardation, and for the neuropsychological cases described in the literature.