A new technique for developing k-distributions is described. This new technique requires almost half of the computational effort yet is at least as accurate as other published k-distribution methods. Its novelty consists in the use of real atmospheric flux calculations to guide both the position of spectral bands and the k terms within each band; previous methods have not used such a guiding principle. Wave number subintervals which have similar atmospheric absorption behavior are chosen, then a representative absorption coefficient for such a wave number interval is set to the value which best fit the results of line-by-line calculations of fluxes and heating rates. This method of choosing one absorption coefficient to represent a large wave number subinterval contrasts with other published methods and is responsible in large part for the improved computational efficiency. It is worth noting that this new technique works as well in the stratosphere as the troposphere, so it can be applied to the processing of satellite data retrievals as well as in weather and climate forecasting. An example of the application of the new technique to the longwave part of the spectrum is presented. A fast k-distribution model (FKDM) suitable for use in weather and climate prediction has been created using 23 k-distribution terms, which is nearly half as many as other k-distribution models. The molecular species represented in the model are H2O, CO2, O3, N2O, CH4, and CFC-11,12,113. FKDM has been developed and validated using a fast line-by-line model (FLBLM). Validations have covered the tropical, midlatitude summer, midlatitude winter, subarctic summer, subarctic winter standard atmospheres, four atmospheres from the Spectral Radiance Experiment campaign, and one case of real tropical atmosphere. It has been found that the FKDM cooling rate accuracy is as follows: 0.15 and 0.2 K day-1 in troposphere for standard and real atmospheres, respectively, and 0.9 K day-1 in all the cases at altitudes below 70 km. Upward and downward flux errors are below 4 W m-2 (usually 1--2 W m-2) in every case.