In spark-ignition engines, Cycle-to-Cycle Variations (CCV) limit the optimization of engine operation since they induce torque variations and the occurrence of misfire and/or knock. A mean for limiting the related negative impact of CCV on fuel consumption and emissions would be control strategies able to address them. At present, engine simulation codes used for control purposes can only describe CCV linked to variations of gas exchanges in the air loop. CCV of the in-cylinder flow motion cannot be naturally captured by classical quasi-dimensional combustion chamber models. A convenient way to mimic CCV is to impose stochastic distributions of the combustion model parameters. Nevertheless, it is not always clear if these perturbations have physical bases as well as realistic ranges of variation. On the other hand, Large-Eddy Simulation (LES) is a natural tool to access local 3D information (flow motion, composition, thermodynamics etc.) at any time and to predict their cyclic variability. The idea followed in this paper is to exploit multi-cycle LES of a Spark Ignition engine in motored operation to analyze the CCV origins and the physical variables that vary in order to define realistic distributions of these quantities. These distributions are implemented in a 1D engine model to simulate various engine operating conditions. It is shown that this methodology allows distinguishing stable from unstable operations. Such engine models can be useful for control purposes aiming at limiting the negative impacts of CCV.