Field measurements are carried out to study statistical properties of the subgrid-scale (SGS) heat fluxes and SGS dissipation of temperature variance in the atmospheric surface layer, and to evaluate the ability of several SGS models to reproduce these properties. The models considered are the traditional eddy-diffusion model, the nonlinear (gradient) model, and a mixed model that is a linear combination of the other two. High-resolution wind velocity and temperature fields are obtained from arrays of 3D sonic anemometers placed in the surface layer. The basic setup consists of two horizontal parallel arrays (seven sensors in the lower array and five sensors in the upper array) at different heights (2.4 and 2.9 m, respectively). Data from this setup are used to compute the SGS heat flux and dissipation of temperature variance by means of 2D filtering in horizontal planes, invoking Taylor’s hypothesis. Model coefficients are measured from the data by requiring the real and modeled timeaveraged dissipation rates to match. Various other experimental setups that differ mainly in the separation between the sensors are utilized to show that filter size has a considerable effect on the various model coefficients near the ground. For the basic setup, conditional averaging is used to study the relation between large-scale coherent structures (sweeps and ejections) and the SGS quantities. It is found that under unstable conditions, negative SGS dissipation, indicative of backscatter of temperature variance from the subgrid scales to the resolved field, is most important during the onset of ejections transporting relatively warm air upward. Large positive SGS dissipation of temperature variance is associated with the end of ejections (and/or the onset of sweeps) characterized by strong drops in temperature and vertical velocity under unstable conditions. These results are also supported by conditionally sampled 2D (streamwise and vertical) velocity and temperature distributions, obtained using an additional setup consisting of the 12 anemometers placed in a vertical array. The nonlinear and mixed model reproduce the observations better than the eddy-diffusion model.