We present a new Krylov subspace method for solving large, sparse and nonsymmetric systems of linear equations. The method, which we denote by ML(k)BiCGSTAB, is a natural generalization of the popular BiCGSTAB method and is derived from a multiple Lanczos procedure using multiple starting left vectors and one starting right vector. Compared with the original BiCGSTAB, our new method requires fewer matrix-vector products at each iteration step, on average requiring only 1 + 1/k matvec's per step. Mathematically, ML(k)BiCGSTAB can be viewed as a bridge connecting the Lanczos-based BiCGSTAB and the Arnoldi-based FOM/GMRES. Empirically, it also seems to be more stable and faster convergent.