In recent years, the interest in riding in cities using the two-wheeler (e.g., bicycles, electric bicycles, electric mopeds, etc.) increases. Mixed-traffic road segments are one of the most common traffic scenes where the mixed two-wheeler flows exist. Because the movements are often not restricted by lanes, the two-wheeler uses lateral road space more freely and shows obvious multilateral interactions (i.e. multi-interaction) with others, bringing issues that endanger traffic safety. A precise estimation of its impacts on traffic operation and safety is necessary, while the microscopic simulation model can satisfy the need as a helpful tool. However, most existing simulation models of these three types of two-wheelers are essentially focusing on handling the one-on-one interaction. The capability to deal with the two-wheeler multi-interaction in mixed traffic is still rare, and the description of what endogenous tasks are contained by the multi-interaction has also not given by literature. To this end, this paper first defines what the multi-interaction entails on the operational behaviour level, claiming that it contains three intertwined processes, namely a (mental) perception, a (mental) decision, and a physical process. The (mental) perception and decision processes represent the recognition of interactions and the response to traffic conditions, while the physical process refers to the execution of these mental activities. A three-layer simulation framework has then been developed, where each layer sequentially corresponds to one of the operational behaviour tasks. Integrated component models are also proposed in each layer to cover these operational tasks. A Comfort Zone model is hence put forward to dynamically perceive the multiple interactive road users, while a Bayesian network model is developed to deal with the decision-making process under multi-interaction situations. Meanwhile, a behaviour force model is also proposed to capture the non-lane based movements following the selected behaviour and current interaction states. Finally, we face validate the proposed models by the comparison between simulation results and observations obtained from trajectory dataset. Results indicate the model performance matches the observed interaction and motion well.