The third-generation real-time single-particle mass spectrometer (RSMS-3) was deployed from March to December 2002 in Baltimore, Maryland, as part of the Baltimore Particulate Matter Supersite. The site was located to the east of the downtown area and northwest of many local industrial emission sites. RSMS-3 actively sampled and analyzed over 380,000 individual particles within the 48--770 nm size range. The resulting positive and negative ion spectra for each particle were classified using a neural network algorithm, adaptive resonance theory ART 2-a. A subset of these data, particles analyzed between 1 April and 30 November 2003, is presented here. Over 99% of these particles could be described by 10 major composition types. Ambient number concentrations were determined for each type and correlated with particle size, wind direction, and time of day/year. On the basis of this information, local and regional sources of different composition classes are postulated. Almost 40% of all particles in the Baltimore aerosol are internally mixed, consisting primarily of organic carbon, ammonium nitrate, and ammonium sulfate. Most of these particles are likely derived from regional sources. The remaining particles appear to be derived mainly from local sources and processes and include elemental carbon (almost 30%), ammonium nitrate (over 10%), and various metals (over 20%). The particle composition types found in Baltimore aerosol are compared to previous measurements in Houston and Atlanta.