Small-scale random fluctuations of atmospheric variables are ubiquitous dynamical components in the stable, free atmosphere. There, within the O(1–10 m) vertical wavelength band, spectra of temperature and horizontal velocity often follow either a m−5/3 or a m−3 power law, m being the vertical wavenumber. Using high-resolution vertical profiles obtained by balloon-born instrumentation in the troposphere and stratosphere, we determine experimental probability density functions (PDFs) of velocity and temperature fluctuations in the spectral band (2–20 m) within atmospheric layers which follow one or the other spectral law. PDFs of such band-filtered fluctuations of temperature and velocities (horizontal and vertical) are estimated within 101 seemingly homogeneous atmospheric layers. It appears that PDFs of horizontal velocity fluctuations, once normalized by their r.m.s. values, do collapse towards two significantly different regimes depending upon the spectral law followed in the wavelength band considered. On the other hand, temperature fluctuation PDFs are shown to be close to each other in both regimes. All these PDFs show close-to-exponential tails. Their high kurtosis appears to be mainly related to intermittency of the fluctuations fields, though marginal influence of residual inhomogeneity of the selected layers may be suspected. These results are compared with published results of laboratory and numerical experiments. We wish to emphasize the unexpected non-Gaussian character of these PDFs.