Active learning (AL) is a popular training strategy that involves iteratively selecting training examples for annotation, typically those for which the current model is most uncertain. This allows to learn more effectively with fewer labeled examples beyond what could be achieved with random samples. However, AL research literature often over-simplifies the evaluation by assuming that the model to guide training data acquisition and the successor model used in the final deployment are identical. In real-world scenarios that is almost never the case since 1) due to performance reasons acquisition models often have less complexity compared to the successor models and 2) successor models are frequently replaced by potentially better performing models in productive environments. In this paper, we systematically study the effects of transferring an actively sampled training data set from an acquisition model to different successor models for biomedical image classification tasks. Our research shows that training a successor model with an actively-acquired data set is most promising, if acquisition and successor model are of similar architecture.