NVIDIA releases five open Nemotron 3.5 models for document search and parsing

NVIDIA released three Nemotron 3.5 Embed variants, an 8B Rerank model and a 631M-parameter Parse model, with open weights available for commercial use. The company says adding Rerank to Embed cut agent inference tokens per query by about 13%–14% in its tests while improving retrieval accuracy. NVIDIA says Parse ranked first on ParseBench among models with fewer than 1 billion parameters as of October 20, 2026.