This study is basically conducted in order to model large earthquakes occurred during 1897-1997 in Bangladesh. On account of inherently memory less property, Poisson process may not be appropriate for large earthquakes. As non-Poisson models take into occurrence history, we choose Markov renewal process which is able to emulate a broad range of temporary behavior of large earthquakes and is expected to capture the main characteristics sufficiently, maintaining simplicity in analysis. Both Weibull distribution and a mixture of two exponential distributions are proposed for the inter-event times considering different events and their parameters are estimated using the likelihood method. Finally, we have predicted the one state of next earthquake being known the state of last event & the time passed by the last occurred earthquake.