NVIDIA releases Nemotron 3.5 Embed, Rerank, and Parse

NVIDIA's 21 October 2026 post adds open, commercially usable models to Nemotron Retriever. The set is Nemotron 3.5 Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16, and Parse at 631 million parameters. Embed searches document text and images, including tables, charts, infographics, datasheets, and schematics. Rerank scores a query against each candidate. Parse turns a page image into Markdown, HTML tables, and structured blocks, and can point extracted evidence back at the page.

NVIDIA says Embed 8B BF16 scores 61.17 nDCG@10 on ViDoRe v3.1 and ranked first among dense retrieval models on that leaderboard as of 20 October 2026. The 1B model, distilled from the 8B, scores 58.38. NVIDIA says text-only embeddings stay in line with the previous Nemotron 3 Embed models, so an existing text index can be reused. In NVIDIA's agent tests, adding Rerank 8B cut inference tokens per query by about 14 percent with Embed 8B. Parse, NVIDIA says, ranks first on ParseBench among models under 1 billion parameters, at 75.20 in a triple pass and 10.5 pages a second on one H100. A single pass scores 72.74 at 16.1 pages a second.

Weights and NIM builds are on Hugging Face and build.nvidia.com. NVIDIA also lists Baseten, DeepInfra, and FriendliAI as hosts for some of the models. The same day's NVIDIA post on open models (opens in a new tab) points at the retriever models as available now. Nemotron 3.5 Super VL and Bolt were already covered on 15 October. This release is the retrieval set.

Who it's for: someone building an agent that has to search long documents, including pages that are mostly tables and figures.

Fits: Best for Research stays Perplexity, the product you open for a sourced answer. These models are the retrieval layer under an agent. The research verdict is unchanged.