This paper uses a model which calculates the flame kernel formation and its early development in spark ignition engines to examine the causes of cycle-to-cycle combustion variations. The model takes into account the primary physical factors influencing flame development. The spark-generated flame kernel size and temperature required to initialize the computation are completely determined by the breakdown energy and the heat conduction from burned region to unburned region. In order to verify the model, the computation results are compared with high-speed Schlieren photography flame development data from an operating spark-ignition engine; they match remarkably well with each other at all test conditions.For the application of this model to the study of cycle-to-cycle variation of the early stage of combustion, additional input is required. First, results from an experiment with a fiber optic spark plug are used to estimate the convective motion of the flame kernel in the vicinity of the spark discharge center at the moment of spark discharge. Then, based on each cycle's convective motion, the mean flame kernel radii are calculated for a fixed time after spark for every cycle, and the flame expansion speeds are obtained. It is found that if only the variation in convection motion is taken into account, the standard deviation of the predicted flame expansion speed distribution is much smaller than that determined experimentally. The dispersion in flame kernel development depends on more than the variation in the convection motion. By assuming that the turbulent kinetic energy is a constant fraction of the total charge kinetic energy from cycle to cycle, the calculated results for the cyclic dispersion in the expansion speed are brought into good agreement with experiments. This indicates that turbulence variations play a much larger role in the cycle-to-cycle variations in flame development than do flame kernel convection variations.