Mouse models of human cancer are vital to our understanding of the neoplastic process, and to advances in both basic and clinical research. Indeed, models of many of the major human tumours are now available and are subject to constant revision to more faithfully recapitulate human disease. Despite these advances, it is important to recognize that limitations do exist to the current range of models. The principal approach to modelling has relied upon the use of constitutive gene knockouts, which can often result in embryonic lethality, can potentially be affected by developmental compensation, and which do not mimic the sporadic development of a tumour expanding from a single cell in an otherwise normal environment. Furthermore, simple knockouts are usually designed to lead to loss of protein function, whereas a subset of cancer-causing mutations clearly results in gain of function. These drawbacks are well recognized and this review describes some of the approaches used to address these issues. Key amongst these is the development of conditional alleles that precisely mimic the mutations found in vivo, and which can be spatially and tissue-specifically controlled using 'smart' systems such as the tetracycline system and Cre-Lox technology. Examples of genes being manipulated in this way include Ki-Ras, Myc, and p53. These new developments in modelling mean that any mutant allele can potentially be turned on or off, or over- or under-expressed, in any tissue at any stage of the life-cycle of the mouse. This will no doubt lead to ever more accurate and powerful mouse models to dissect the genetic pathways that lead to cancer.