We propose a general particle filtering and learning framework for mixed-frequency state-space models. Our mixed-frequency particle methods use a smoother so as to draw the Bayesian inference from low-frequency observations. Our forward smoother is simple and efficient, and the sample path degeneracy is negligible with a small lag size. The backward smoother mitigates the sample path degeneracy effect with quadratic computations that are nevertheless parallelizable. To illustrate our mixed-frequency particle framework, we take the mixed-frequency conditional dynamic linear model with regime switching as an example. In a simulation study, we show that naive treatments of mixed frequencies may severely impact model identification.