The purpose of the present paper is to analyze L2 and L3 production and comprehension from a cognitive-pragmatic point of view, taking into account Relevance Theory (Sperber and Wilson, 1986; Wilson and Sperber, 2006), Mental Models Theory (Johnson-Laird, 1983) and the Graded Salience Hypothesis (Giora, 1997). Special attention is paid to error analysis and to the detection of error sources, especially in the case of errors not attributable to transfer, interference or overgeneralization. The paper is based on three studies involving, first, L2 and L3 production (Study 1), both production and comprehension (Study 2) and L3 comprehension (Study 3). In general, the phenomena observed can be explained by a combination of Relevance Theory, Mental Models Theory and the Graded Salience Hypothesis. In fact, even when transfer is used as a strategy, its use is relevant to the learner, who assumes that it will be relevant to the recipient as well. The results also shed some light on the multilingual mental lexicon and multiple language processing.