In textual modeling, models are created through an intermediate parsing step which maps textual representations to abstract model structures. Therefore, the identify of elements is not stable across different versions of the same model. Existing model differencing algorithms, therefore, cannot be applied directly because they need to identify model elements across versions. In this paper we present Textual Model Diff (TMDIFF), a technique to support model differencing for textual languages. TMDIFF requires origin tracking during text-to-model mapping to trace model elements back to the symbolic names that define them in the textual representation. Based on textual alignment of those names, TMDIFF can then determine which elements are the same across revisions, and which are added or removed. As a result, TMDIFF brings the benefits of model differencing to textual languages.