This paper studies identification and estimation of discrete games in large networks, with an application to peer effects on smoking in friend networks. Due to the presence of multiple equilibria, the model is not point identified. We adopt the partial identification approach by constructing moment inequalities on choice probabilities of subnetworks. Doing so not only significantly reduces the computational cost, but also enables us to find consistent estimator of the moment conditions even when the network is large and the friendship relationship structure varies significantly among networks. In the application using the Add Health data, we find significant and positive peer effects on smoking.