Simulation of human blood flow is a demanding task both in terms of the complexity of applicable models and the computational effort. One reason is the particulate nature of blood which in first approximation may be treated as a suspension of red blood cells (RBCs) in blood plasma. A second reason is that in realistic geometries typical length scales vary over several orders of magnitude. Usual computational models either cope with this complexity by implementing only a homogenous, although non-Newtonian, fluid or resolve with high accuracy a relatively small numbers of RBCs by means of deformable meshes. We present a coarse-grained and highly-efficient yet particle-based model for blood that allows us to simulate up to millions of cells on current parallel supercomputers. We start with a Lattice Boltzmann based method for the simulation of suspensions of rigid particles which accounts for long-range hydrodynamic interactions. Real RBCs, however, are not rigid. We thus add anisotropic model potentials to cover the more complex short-range behavior of deformable cells on a phenomenological level. We will discuss the effect of the model parameters and demonstrate the applicability of the model to simple situations of confined flow as well as its scalability.