<p>Northwest China (NWC), situated in an arid–semi-arid climatic setting, is highly sensitive to global warming and shows pronounced changes in temperature extremes. Using NEX-GDDP-CMIP6 and the CN05.1 gridded observations, we evaluate model performance for eight temperature-extreme indices (FD, ID, CSDI, TNn, TR, SU, WSDI, and TXx) and an additional absolute-threshold heat index, T35 (days with T_max &gt; 35 °C), during 1961–2014. Based on the evaluations of spatial distributions from climatological mean and linear trends, we select a subset of models with better agreement with observations to estimate how the temperature-extreme indices may evolve in the future for SSP3-7.0 and SSP5-8.5. The evaluation results show that the multi-model ensemble reproduces the observed spatial distributions reasonably well, especially for frequency indices, and reveals an asymmetric historical warming pattern with fewer cold extremes and more warm extremes. Projections show widespread declines for FD, along with reductions in ID, and strong upward shifts in TR, SU, and T35, with larger amplitudes under SSP5-8.5. The strongest rise in warm nights occurs over the Tarim Basin, while the decline in cold-event frequency is comparatively moderate in high-elevation regions such as the northern Kunlun Mountains. Return-period analysis shows that the occurrence probability for representative cold-extreme events declines markedly under future scenarios. In contrast, heat extremes with a historical return period of ~ 3000 years are anticipated to occur far more often, with return period shortening to about 10 years under SSP3-7.0 and approaching near-annual recurrence under SSP5-8.5 scenarios. This quantitatively demonstrates a rapid shift in extreme-temperature risk from cold extremes toward heat extremes over NWC. Overall, our results quantify contrasting responses of temperature extremes in arid NWC and provide evidence for understanding extreme warming and heat-risk assessment.</p>