Large-scale inversion of gravity gradient tensor data is a time-consuming problem with high demands on computational and physical memory usage. To avoid extraordinary matrix-vector multiplications in each inverse iteration and to speed up the forward of geophysical models, planting inversion is introduced and conjugate gradient iteration replaced by accumulation summary based on L1 norm. The planting inversion easily leads to adjacent anomalies mutually invasive, a horizontal weighted function is proposed to suppress mutual interference between the adjacent anomaly sources. These results of the inversions and analysis results show that planting inversion with horizontal weighted function obtain a meaningful geophysical model. And these methods require little memory and high efficiency.