Abstract: Based on the TIGGE data, the ensemble mean outcomes of the 24-168 hours ensemble forecasts for the surface temperature in the Northern Hemisphere during the summer of 2007 are provided by ECMWF, JMA, NCEP and UKMO. Root-mean-square errors (RMSE) of the results are examined.The multi-model ensemble forecasts of the surface temperature for the forecast period from August 1 to 31 of 2007 are conducted by utilizing the superensemble, the multi-model ensemblemean, and the bias-removed ensemble mean.The forecast skills of these multi-model ensemble methods are investigated.The results show that the JMA model performs best for 24-hour and 48-hour forecasts, but for 72-hour to 168-hour forecasts the ECM WF model is the one with the highest skill. Forecast skills of these four models of different operational forecast centers are quite different over various regions.The JMA model performs best in the United States and Europe, while in China the UKMO model is the best one. Actually, none of the four models is perfect for all regions in the Northern Hemisphere.Therefore, it is necessary to utilize the multi-model superensemble methodology to improve the forecast skill.The multi-model superensemble and the bias-removed ensemble mean reduce the RMSE of the surface temperature forecast considerably.Both methods show an improvement on the forecast skill over the best single model forecast and the multimodel ensemblemean.In early (late) phase of the forecast period, the multi-model superensemble produces a smaller (larger) RMSE than the bias-removed ensemble mean for the 24-hour and 48-hour forecasts.For the 72-hour to 168-hour forecasts, both the multi-model superensemble and the the bias-removed ensemble mean lead to smaller RMSE than the multi-model ensemble mean and single models during the first two weeks of the forecast period.Never theless, two weeks later the RMSE of the superensemble forecast increases rapidly, and exceeds that of the bias-removed ensemble mean.The RMSE of the multi-model superensemble forecast may even exceed that of the single model forecast at the end of the forecast period.Weight coefficients of multi-model superensemble methodology remain unchanged during the fixed training period. It's expected that an improved multi-model superensemble methodology with flexible training period when statistical weights of the multi-models vary with time may further improve the forecast skill.