Abstract: Using the 3-h national site observation data at Beijing area in spring (February to April) during 2019 to 2021, refined evaluation of temperature forecasts by current main-stream numerical models (ECMWF and CMA_GFS), national gridded guidance forecast product (SCMOC) and provincial and municipal revised feedback gridded forecast product (SMERGE) are conducted. The results show that, the temperature forecasts for spring in Beijing often show negative bias in the models of ECMWF and CMA_GFS. This bias has no significant difference between mountainous and plain areas, but is more prominent during nighttime periods. The gridded forecast products (SCMOC and SMERGE) have a good ability to correct the temperature forecasted by the models. The temperature forecast biases of the gridded products are concentrated between -1 to 1℃, and the forecast accuracy is higher and the mean absolute error is lower than that of model forecast. There are issues with the forecasts of 24-h temperature change and daily temperature range for the four products. The amplitude of strong 24-h temperature change forecasted by all products if relatively smaller than that of observation, and the gridded forecast products do not demonstrate significant correction ability. In addition, the daily temperature range forecasted by all products has a positive deviation of 1~3℃ compared to the observation. SCMOC has better correction ability for the daily temperature range forecasted by the models, while SMERGE overestimates more prominently than that in models. The positive deviation of daily temperature range forecast is closely related to the underestimation of low temperature (at 05:00 BT) in the model, while the overestimation of high temperature (at 14:00 BT) can not be ignored in the grid forecast. It is suggested that gridded forecast products should not only focus on improving (reducing) overall accuracy (biases), but also on the improvement of the forecasts of development and evolution of important synoptic processes.