This study proposes a temperature model for the relaxation of magnetic nanoparticles and a phase measurement method under a mixing-frequency excitation field, which can improve the accuracy of temperature measurements in magnetic nanothermometry. According to the Debye-based magnetization model for magnetic nanoparticles, phases at mixing frequencies are used to solve the problem of a delay in the relaxation phase of the magnetic field at a high frequency. This method can improve the signal-to-noise ratio of the response of the magnetic nanoparticles and weaken the phase shift of the detection coils caused by the changes in temperature. The results of experiments show that the proposed method can achieve static temperature measurement error less than 0.1 K and dynamic temperature measurement error less than 0.2 K.