We propose a new kernel estimator of conditional density and derive its asymptotic bias and variance. This new, non-negative estimator, is obtained by ‘internalizing’ the random denominator of the well known local constant smoother of Rosenblatt (1969). We analyze the performance of the new procedure, and that of various competitors in a variety of Monte Carlo experiments. We also illustrate the applicability of the new estimator using S&P 500 options prices to study implied volatility.