The mesoscale prediction of winds and precipitation over the Northeast U.S. is often complicated by the Appalachian terrain, coastline curvature, and various urban centers. Through a collaborative effort between Stony Brook University and the NWS operational forecasters over the Northeast, the Penn State-NCAR Mesoscale Model (MM5) has been run in real-time during the past four years down to 4-km grid spacing for the southern New England and New York City Metropolitan region (http://atmos.msrc.sunysb.edu/mm5rt). This effort has improved our understanding of the model skill at high resolution and the impact air-sea interactions have on the forecasts. Several model biases have been documented such as the warm bias over water during the cool season, an urban cool, moist, and weak wind bias during the day, and a dry bias during the warm season. These biases impact the coastal mesoscale flows. For example, 4-km sea breezes are 1-2 hours early on average, and there is generally too little warm season explicit precipitation in the high resolution nests. More observations are needed over the coastal waters to verify mesoscale models. To meet this challenge, Stony Brook in collaboration with the NWS has instrumented a passenger/car ferry across Long Island Sound between Bridgeport, CT and Port Jefferson, NY to collect real-time winds, temperatures, moisture, and surface flux measurements. The 4-km MM5 forecasts are now being verified using this unique dataset, which have helped validate MM5 sea breeze circulations over the Sound as well as quantify problems with MM5 temperatures, winds, and fluxes over water. In order to help improve the mesoscale forecasts over the Northeast, a mesoscale ensemble forecast system has been developed at Stony Brook (http://fractus.msrc.sunysb.edu/mm5rte). During the warm season of 2003, the MM5 for the 0000 UTC cycle was run down to 12-km grid spacing using 12-different physics parameterizations (3 different boundary layer and 4 different convective parameterizations) and 7 different initial conditions (Eta 0000 UTC, NCEP Eta breds, and GFS). This presentation will highlight the benefits and weaknesses of such an ensemble system for forecasting winds and precipitation over this region, and discuss the differences between the physics-based versus initialization-based ensemble.