ULTRA is a basic model for knowledge graph reasoning. A single pre-trained ULTRA model can execute link prediction tasks on any multi-relational graph and support any entity/relationship vocabulary. Its performance surpasses many SOTA models that are trained specifically for each graph. Following the pre-training-fine-tuning paradigm of basic models, pre-trained ULTRA checkpoints can be immediately used for zero-shot reasoning on any graph, as well as further fine-tuning. ULTRA provides a unified, learnable, and transferable representation for any knowledge graph. ULTRA employs graph neural networks and a modified version of NBFNet. It does not learn specific entity and relationship embeddings for the downstream graph, but rather acquires relative relationship representations based on the interaction between relationships.