In order to achieve multi-input-multi-output (MIMO) non-Gaussian vibration test control, an inverse multi-step prediction method for MIMO non-Gaussian vibration is proposed. This method constructs a zero memory nonlinear transformation based on the generalized Gaussian distribution, and generates non-Gaussian reference signals from the transformation. The inverse multi-step prediction model of the system is derived, and then the reference signals are used as the inputs of the model to generate the driving signals. In the control process, the reference non Gaussian signals are adjusted according to the error between estimated power spectral density and kurtosis of the control point and their reference values. And the driving signals are updated by using the inverse multi-step prediction model, thus achieving MIMO non Gaussian vibration control. Finally, a three-axis shaking table test is carried out. The test results show that the proposed method has good control results in both the power spectrum and kurtosis, thus its feasibility is verified.