Soil carbon, nitrogen, and phosphorus are crucial nutrients supporting the soil quality and vegetation growth of temperate meadow steppes. Estimating these nutrients using hyperspectral data is significant for rapidly and accurately monitoring soil nutrient information in temperate meadow steppes. This study focused on grasslands with three different utilization patterns (grazing, mowing, and enclosure) in the Hulunbeir temperate meadow steppe. Eighteen sample plots were selected, with three replicates per plot, and soil samples from a depth of 0 to 30 cm were collected to measure soil carbon, nitrogen, and phosphorus contents as well as hyperspectral data. Models for inverting soil carbon, nitrogen, and phosphorus stoichiometric ratios using hyperspectral data were established through Back Propagation Neural Network (BPNN), Random Forest (RF), and Partial Least Squares Regression (PLSR). The optimal model was selected by comparing R2 and Root Mean Square of Residuals (RMSR). The results showed that: (1) The RF model performed excellently in spectral inversion of total carbon (TC), total nitrogen (TN), and total phosphorus (TP) contents under the three utilization patterns (R2 ≥ 0.4433, RMSE ≤ 12.0604), followed by the BPNN model. PLSR was only applicable to the inversion of TC, TN, and TP contents under grazing; (2) Under grazing, all three models performed well in spectral inversion of soil carbon-to-nitrogen (C/N), carbon-to-phosphorus (C/P), and nitrogen-to-phosphorus (N/P) ratios (R2 ≥ 0.4144, RMSE ≤ 65.4081); (3) Under mowing, BPNN performed well in C/P spectral inversion (R2 = 0.9916, RMSE = 7.0938), followed by PLSR. Only RF performed well in C/N spectral inversion (R2 = 0.7749, RMSE = 0.3028); (4) Under enclosure, all three models performed well in spectral inversion of C/N and C/P (R2 ≥ 0.4462, RMSE ≤ 24.0289), while only PLSR performed well in N/P spectral inversion (R2 ≥ 0.7172, RMSE ≤ 0.8171). Overall, this study suggests that the RF model has stronger applicability in the Hulunbeir temperate meadow steppe region. The findings aim to provide theoretical and technical support for quantitative inversion of surface soil carbon, nitrogen, and phosphorus in temperate meadow steppes with different utilization patterns based on hyperspectral inversion.