We study the physical properties of 216 z similar or equal to 2.1 Ly alpha-emitting galaxies (LAEs) discovered in an ultra-deep narrow-MUSYC image of the ECDF-S. We fit their stacked spectral energy distribution (SED) using Charlot & Bruzual templates. We consider star formation histories (SFHs) parameterized by the e-folding time parameter tau, allowing for exponentially decreasing (tau > 0), exponentially increasing (tau < 0), and constant star formation rates (SFRs). We estimated the average flux at 5015 angstrom of our LAE sample, finding a non-detection, which translates into negligible He II line emission at z similar or equal to 2.1. In addition to this, the lack of high equivalent width (EW) Ly alpha line objects ruled out the hypothesis of a top-heavy initial mass function in LAEs. The typical LAEs of our sample are characterized by best-fit parameters and 68% confidence intervals of log(M(*)/M(circle dot)) = 8.6[8.4-9.1], E(B - V) = 0.22[0.00-0.31], tau = -0.02[(-4)-18] Gyr, and age(SF) = 0.018[0.009-3] Gyr. Thus, we obtain robust measurements of low stellar mass and dust content, but we cannot place meaningful constraints on the age or SFH of the LAEs. We also calculate the instantaneous SFR to be 35[0.003-170] M(circle dot) yr(-1), with its average over the last 100 Myr before observation giving < SFR >(100) = 4[2-30] M(circle dot) yr(-1). When we compare the results for the same SFH, typical LAEs at z similar or equal to 2.1 appear dustier and show higher instantaneous SFRs than z similar or equal to 3.1 LAEs, while the observed stellar masses of the two samples seem consistent. Because the majority are low-mass galaxies, our typical LAEs appear to occupy the low-mass end of the distribution of star-forming galaxies at z similar or equal to 2. We perform SED fitting on several sub-samples selected based on photometric properties and find that LAE sub-samples at z similar or equal to 2.1 exhibit heterogeneous properties. The typical IRAC-bright, UV-bright, and red LAEs have the largest stellar mass and dust reddening. The typical UV-faint, IRAC-faint, and high EW LAE sub-samples appear less massive (<10(9) M(circle dot)) and less dusty, with E(B - V) consistent with zero.