In four experiments we explored processes underlying L2 gender retrieval. We focused on L1 interference and on the influence of the L2 noun's termination. In Experiments 1 and 2 we tried to manipulate the intensity of L1 interference. We found that L2 speakers cannot eliminate or substantially reduce the interlingual interference neither when they know the response language long in advance in a situation in which code-switching is required (Experiment 1), nor when they are close to the monolingual mode (Experiment 2). Experiments 3 and 4 yielded evidence that gender typicality of the L2 noun's termination also exerts an influence on L2 processing, both in production and comprehension. L2 gender thus does not seem to be stored as a fixed feature as it is assumed for L1. Rather, our data support the assumption that it is computed anew each time when needed for processing. Further implications for modeling are discussed.