NVIDIA's Nemotron 3.5 Embed, Rerank and Parse put open retrieval models under a permissive license

Most retrieval pipelines still read a PDF by pulling out its text layer and hoping the tables survive. NVIDIA’s new Nemotron Retriever models, published on 21 October alongside its GTC Berlin keynote, go after that step directly: they embed whole page images, rerank what comes back, and parse pages into structured Markdown.

What shipped

Model	Size	Job
Nemotron 3.5 Embed 8B BF16	8B	Embeds text, page images, or both
Nemotron 3.5 Embed 1B BF16 / NVFP4	1B	Distilled from the 8B, smaller footprint
Nemotron 3.5 Rerank 8B BF16	8B	Cross-encoder that rescores query and page pairs
Nemotron 3.5 Parse	631M	Turns page images into Markdown, HTML tables and blocks

The Embed 8B card lists a 32,768-token context and a 4096-dimension vector that can be truncated to 1024 or 512 dimensions, as long as you re-normalize it. Queries are text only; documents can be an image, text, or both. The card shows Transformers, Sentence Transformers and vLLM recipes.

The license

All five cards point to the OpenMDW-1.1 license and say the models are “ready for commercial or non-commercial use”. None is gated. OpenMDW grants rights “without restriction”, asks you to keep the license and notices when you redistribute, ends your rights if you sue claiming the models infringe, and puts no restrictions on outputs.

What NVIDIA reports

These are NVIDIA’s numbers, not ours:

• ViDoRe v3.1: Embed 8B at 61.17 nDCG@10, first among dense retrievers as of 20 October; Embed 1B at 58.38.
• Reranking: 60.49 against 59.17 for Qwen3-VL-Reranker-8B on a new ViDoRe v3.1 reranking board built with the MTEB team.
• Agent tokens: adding the reranker cut agent tokens per query by about 14% with Embed 8B and 13% with Embed 1B, with the agent and tasks held fixed.
• Parse: 75.20 on ParseBench at 10.5 pages per second on one H100 in triple-pass mode, or 72.74 at 16.1 pages per second in a single pass.

Before you swap

NVIDIA says the language half of the model stayed frozen in training, so text-only embeddings should match the older Nemotron 3 Embed models and existing text indexes can be reused. That is worth one quick cosine check on your own data before you rely on it. Parse is a little older than the announcement: its card gives a 17 October release date. Weights are on Hugging Face, with hosted options from NIM, Baseten, DeepInfra and FriendliAI.