A multimodel ensemble air quality forecasting system was created as part of the New England Air Quality Study (NEAQS-2004) during the summer of 2004. Seven different models were used, with their own meteorology, emissions, and chemical mechanisms. In addition, one model was run at two different horizontal grid resolutions, providing a total of eight members for the ensemble. Model forecasts of surface ozone were verified at 342 sites from the EPA's AIRNOW observational network, over a 56 day period in July and August 2004. Because significant biases were found for each of the models, a simple 7-day running mean bias correction technique was implemented. The 7-day bias correction is found to improve the forecast skill of all of the individual models and to work nearly equally well over the entire range of observed ozone values. Also, bias-corrected model skill is found to increase with the length of the bias correction training period, but the increase is gradual, with most of the improvement occurring with only a 1 or 2 day bias correction. Analysis of the ensemble forecasts demonstrates that for a variety of skill measures the ensemble usually has greater skill than each of the individual models, and the ensemble of the bias-corrected models has the highest skill of all. In addition to the higher skill levels, the ensemble also provides potentially useful probabilistic information on the ozone forecasts, which is evaluated using several different techniques.