NVIDIA has refreshed the retrieval side of its open model lineup. In a Hugging Face post dated October 21, 2026, the company introduced Nemotron 3.5 Embed, Nemotron 3.5 Rerank and Nemotron 3.5 Parse, five new open models it says are licensed for commercial use. They join the Nemotron Retriever collection, previously called Nemotron RAG, and are aimed at the search and document-ingestion steps that sit underneath enterprise AI agents.
The embedding family comes in three versions: an 8B model in BF16, a 1B model distilled from it, and the same 1B model quantized to NVIDIA's NVFP4 format for both its text and image encoders. Unlike the text-only Nemotron 3 Embed generation, these models index document images, so tables, charts, infographics, datasheets and schematics become searchable alongside ordinary text. NVIDIA built the 8B embedder on Ministral-3-8B-Instruct and kept the language component frozen during training, which it says leaves text embeddings unchanged and lets customers reuse existing vector indexes.
Rerank 8B is a cross-encoder on the same backbone. Rather than comparing separately computed vectors, it reads each query and candidate document together and outputs a relevance score. Parse, at 631 million parameters, turns page images into Markdown, HTML tables and structured blocks while preserving reading order and the position of each element, so extracted evidence can be traced back to its spot on the original page.
NVIDIA's own benchmark results are strong. By its account, Embed 8B scores 61.17 nDCG@10 on ViDoRe v3.1, first among dense retrievers and nearly nine points above its previous visual embedder, while the distilled 1B model keeps about 95% of that score. Rerank 8B posts 60.49 on a new ViDoRe v3.1 reranking leaderboard launched with the MTEB team, ahead of Qwen-3-VL-Reranker-8B at 59.17. Parse leads ParseBench among models under one billion parameters with 75.20 in its slowest, three-pass mode, processing 10.5 pages per second on a single H100; a single pass reaches 16.1 pages per second at a lower score.
The agent angle is about cost as much as accuracy. In NVIDIA's tests on ViDoRe v3, BRIGHT and BrowseComp-Plus, adding the reranker cut the inference tokens an agent spent per query by roughly 14% with the 8B embedder and 13% with the 1B version, because the agent needed fewer repeated searches. The NVFP4 variant raised throughput on GB200 by about 27% while keeping nearly all of its accuracy, and NVIDIA's NIM serving stack handled about 2.9 times the queries of a comparable vLLM setup on an RTX PRO 6000 Blackwell card.
Weights, training recipes and fine-tuning tools are on Hugging Face, with NIM containers and hosting through Baseten, DeepInfra and FriendliAI. Partners including Elastic, Perplexity, Vectorize and Turbopuffer report favorable early evaluations, and NVIDIA's separate Agent Toolkit announcement the same day lists Amazon OpenSearch Service, SAP, ServiceNow and Zoom among users of Nemotron Retriever.
The caveats are familiar. Leaderboard positions are dated October 20, 2026 and can change quickly, the head-to-head comparisons and agent token savings come from NVIDIA's own experiments, and partner results were run on private or vendor-chosen benchmarks. The models ship under NVIDIA's own license terms, which teams should read before deployment.