We have implemented a model-generation based parallel theorem prover in KL1 on a parallel inference machine, PIM. We have developed several techniques to improve the efficiency of forward reasoning theorem provers based on lazy model generation. The tasks of the model-generation based prover are the generation and testing of atoms to be the elements of a model for the given theorem. The problem with this method is the explosion in the number of generated atoms and in the computational cost in time and space, incurred by the generation processes. Lazy model generation is a new method that avoids the generation of unnecessary atoms that are irrelevant to obtaining proofs, and to provide flexible control for the efficient use of resources in a parallel environment. With this method we have achieved a more than one-hundred-fold speedup on a PIM consisting of 128 PEs.