Abstract: We develop an automatic centroid moment tensor (CMT) inversion system for large earthquakes (M_W≥6.5) at global scale based on the W-phase method. We use seismic waveforms observed by global virtual seismic networks including Chinese stations, which have been recently set up. To speed up the calculation of moment tensor solutions, the W-phase method is parallelized using openMP. In order to assess the accuracy and timeliness of this implementation, automatic moment tensor solutions for 140 global events (M_W6.5—9.0) are determined offline. Compared with the global CMT (GCMT) solutions, the result indicates that this system mentioned is able to precisely retrieve CMT solutions of global events with M_W≥6.5. There is remarkably good linear relationship between the moment magnitude of W-phase CMT and GCMT. The standard deviation of the magnitude difference ΔM_W is about 0.13, and 96 percent of the inversions yield a M_W within M_W-GCMT from the GCMT ±0.2. In comparison with the GCMT solutions, six elements of moment tensors derived show nearly linear distribution along the diagonal, respectively. 84 percent of inversions yield a great-circle distance smaller than 50 km between the CMT and GCMT centroids. As shown above, majority of the centroids are very similar with the GCMT centroids. When the data coverage is sufficient, this system can thus automatically determine the CMT of global earthquakes with M_W≥6.5 in between 25 min and 40 min after the origin time.