We propose an offline goal-conditioned RL algorithm that achieves state-of-the-art performance on complex, long-horizon tasks without needing hierarchical policies or generative subgoal models.