NVIDIA Expands Nemotron Retriever With Models Aimed at Cutting Agent Search Tokens

NVIDIA has expanded its Nemotron Retriever collection with new models for multimodal search, reranking, and document parsing, reporting that reranking cut agent inference tokens per query by about 13% to 14% in its evaluations. The models are designed to help agents find relevant evidence across text and visually rich documents, potentially reducing repeated searches before an agent can answer. (NVIDIA’s model overview)