The maximum freezing depth is an important indicator for the thermal state of seasonally frozen ground， and its changes have an important impact on the regional water cycle， ecological processes and engineering stability. This paper released a soil maximum freezing depth grid dataset for 10-year period from 1961 to 2020 in Northwest China and Tibet， with a spatial resolution of 1 km. The dataset was produced by integrating downscaled and bias corrected weather data， elevation and soil properties using a support vector machine model with 200 ensemble simulations. The 10-fold cross-validation shows that the accuracy of the support vector machine model is acceptable ［R² = 0.70 ± 0.29， RMSE = （23.63 ± 10.30） cm， bias = （-0.77 ± 6.01） cm］. Validation using in-situ data shows that the R² for the four periods 1980s， 1990s， 2000s and 2010s are 0.77， 0.83， 0.73 and 0.71 respectively， and the RMSE are 27.14 cm， 22.42 cm， 21.63 cm and 23.58 cm respectively. The uncertainty of the simulation results is stable throughout the simulation period. Based on this dataset， we found that the soil maximum freezing depth in the Northwest China and Tibet decreased significantly between 1960s and 2020s， with an average rate of 3.02 cm per decade. The dataset can be downloaded via the National Tibetan Plateau/Third Pole Environment Data Center （DOI： 10.11888/Geocry.tpdc.271774）.