In this paper, we address the problem of system-theoretic dynamic information fusion in time-varying heterogeneous sensor networks. This class of sensor networks involves nodes that receive observations from a process of interest (active nodes) and nodes that do not receive any information (passive nodes), where the active and passive node roles can be varying with respect to time. At any given time, in addition, active nodes are allowed to have nonidentical modalities such that they can observe different measurements from the process. Specifically, we propose a new distributed input and state "coestimation" architecture for time-varying heterogeneous sensor networks, where time evolution of input and state updates of each node both depend on the local input and state information exchanges. The stability and performance of the overall sensor network are guaranteed once the local sufficient stability conditions for each node are satisfied. As compared with our recent distributed input and state "estimation" approach for the same problem, where time evolution of input (respectively, state) update of each node only depends on the local input (respectively, state) information exchange, our illustrative numerical example also demonstrates a substantially improved dynamic input and state fusion performance.