This paper provided a multivariate phase space reconstruction method with good anti-noise performance by making use of the third-order cumulant which was used to reflect high-order nonlinear correlation between multivariate series. First, the third-cumulant was introduced to calculate the local intrinsic dimension (LID) of series, and a new third-order cumulant correlation matrix based on different phase points was constructed. An evaluation function was introduced to find an appropriate third-cumulant slice which possessed strong robustness property on the noise and the reconstruction parameters, like embedding dimension. Then, the embedding dimension and the embedding delay were calculated to reconstruct the phase space of multivariate series. Finally, simulation results were given to show that the approach proposed was more effective for noisy chaotic series, from which it could be clearly seen that the multivariate strange attractors got better extensions in the reconstructed phase space.