We propose a new semiparametric observation-driven volatility model which combines the Generalized Autoregressive Score (GAS) model with kernel density estimation. We use the kernel density method to estimate the error density. Unlike semiparametric GARCH models, the form of the error density also governs the specification of volatility dynamics. We provided simulated evidence of estimation efficiency and forecasting accuracy of the new model. We also apply the model to IBM return data and show that the model does a good job of density forecasting.