We present an instance-based nearest neighbor approach to entity linking. In\ncontrast to most prior entity retrieval systems which represent each entity\nwith a single vector, we build a contextualized mention-encoder that learns to\nplace similar mentions of the same entity closer in vector space than mentions\nof different entities. This approach allows all mentions of an entity to serve\nas "class prototypes" as inference involves retrieving from the full set of\nlabeled entity mentions in the training set and applying the nearest mention\nneighbor's entity label. Our model is trained on a large multilingual corpus of\nmention pairs derived from Wikipedia hyperlinks, and performs nearest neighbor\ninference on an index of 700 million mentions. It is simpler to train, gives\nmore interpretable predictions, and outperforms all other systems on two\nmultilingual entity linking benchmarks.\n