In this talk, we will share our experience of implementing a Peer-learning system, PeerWise at University College London. PeerWise is an online repository of multiple-choice questions that are created, answered, rated and discussed by students. PeerWise involves students in the formative assessment and feedback process and enables them to develop a number of key skills which will enhance the employability of our students, including negotiating meaning with others, cross-cultural communication, and analytical and evaluation skills as they engage with the work of their peers. The system was implemented with one linguistic module for undergraduate and postgraduate students. Students were put into groups of three. The method of grouping was manipulated to be either randomly or by mixed ability. Firstly, we found that there is an active engagement throughout the module, not just meeting deadlines. Furthermore, there was an increase in usage before exams which suggests that it was used as a revision tool. Secondly, we found an significant correlations between their module performance and PeerWise internal scores which is composed of individual scores for question writing, answering questions and rating existing questions. This only applied to students that were grouped by mixed ability. Finally, we performed a nested model comparison to test if the mixed ability grouping has an effect on their performance. The lm function in R statistical package was used to build two models, the superset model which has both PeerWise score and Group as the predictors, and the subset model which has only the PeerWise score as the predictor. It was found that while both PeerWise scores and grouping were significant predictors in the superset model, grouping made a significant improvement in prediction with p < 0.05. Together, our results suggest that mixed-ability grouping is key to peer learning.