Non-stationary probability models of extreme values following generalized extreme value (GEV) distributions were fitted to annual climate extremes of maximum temperature, minimum temperature, and maximum daily precipitation from 489 observational stations in China spanning from their own initial times to 2013. The non-stationary return level of extremes and its derivative with respect to time as functions of return period and time were also derived. Trends and spatial patterns of linear changing rates of "ordinary state" (2-year-period return level, denoted as z_0.5) and "extreme state" (50-year-period return level, denoted as z_0.02) of climate extremes were analyzed. A special case of opposite tendencies for z_0.5 and z_0.02 was particularly investigated for its origin and possible impact. The results show that, for maximum temperature, stationary models generally fit for the monsoon area of East China, while non-stationary models dominate for other regions in China and most of them indicate increasing trends for both z_0.5 and z_0.02. For minimum temperature, non-stationary models fit nationwide in China and also indicate increasing trends for both z_0.5 and z_0.02 in general, except for a part of Northeast China where z_0.02 shows a decreasing trend opposite to z_0.5. For maximum daily precipitation, stationary models fit overwhelmingly in China. When the scale parameter of GEV distribution changes with time, the variability of z_0.02 is much greater than that of z_0.5 and, consequently, opposite tendencies may happen for them. In particular, when z_0.5 becomes mild as z_0.02 becomes more extreme, much more intense disastrous weather may occur.