This paper deals with the computational simulation of atmospheric cloud dispersion resulting from liquefied gas spills. Experimental data were taken from the Coyote series trials that involved large scale cryogenic natural gas releases. The data were used for checking the validity of the results obtained through the computational fluid dynamics code CFX. Moreover, using the experimental data-set of the non-isothermal gas dispersion, a comparative evaluation was performed between the code and two popular box-models (SLAB and DEGADIS) on the basis of specific statistical performance measures. The results showed that CFX approximated gas concentration histories with a reasonably good agreement and predicted correctly the behavior of gas cloud during dispersion. In particular, it predicted that the cryogenic dense cloud produced by the flash boiling would disperse as a heavy rather than a light gas, although the natural gas is lighter than air at normal conditions. On the other hand, the statistical treatment revealed that box-models may also give acceptable results even for the complex case of non-isothermal gas dispersion. Nevertheless, the CFX code unambiguously demonstrated considerably better accuracy. (c) 2005 Elsevier Ltd. All rights reserved.