NVIDIA's Nemotron 3.5 Embed and Parse: document pages as images

NVIDIA released Nemotron 3.5 Embed, which searches page images and text without OCR, and Nemotron 3.5 Parse, a 631M page parser, both under OpenMDW-1.1 for commercial use.

  • New release
  • 2 min read
  • Published

NVIDIA published two document models on Hugging Face in October 2026: Nemotron 3.5 Parse on 17 October and Nemotron 3.5 Embed on 20 October. NVIDIA announced them on 21 October together with a reranker, Nemotron 3.5 Rerank. One turns a page image into structured text, and the other lets you search page images with an ordinary text question.

What is new

Embed is a retrieval model for document pages. Queries are text, while documents can be a page image, text or both, so tables, charts and schematics can be searched without running OCR first. There is an 8B model (8.38B parameters, 4,096-dimensional vectors, 16.8 GB in BF16), a 1B model distilled from it (1.55B parameters, 2,048 dimensions, 3.1 GB) and an NVFP4 build of the 1B (1.3 GB). The language part is frozen and comes from NVIDIA’s text-only Nemotron 3 Embed models, so text embeddings are identical and existing text indexes can be reused. NVIDIA reports 61.17 nDCG@10 for the 8B model on ViDoRe v3.1, first among dense retrieval models on that leaderboard on 20 October, against 52.23 for its earlier Llama Nemotron Embed VL 1B.

Parse reads a page image into Markdown, tables, chart data, reading order and bounding boxes. It has 631 million parameters. You choose how much work each page gets: a single pass parses the page once, a double pass re-parses tables and charts at higher detail, and a triple pass adds a layout detector. NVIDIA reports ParseBench scores of 72.74, 74.92 and 75.20 for the three modes, at 16.1, 11.2 and 10.5 pages per second on one H100. The triple pass is the top score in NVIDIA’s comparison, but only 0.1 points above LightOnOCR-3-4B at 75.10. All of these figures are NVIDIA’s own.

Running it

Embed’s cards give examples for Transformers, Sentence Transformers and vLLM. Text retrieval was evaluated in 34 languages and image retrieval in 18, and queries must be text, so you cannot search with an image. Parse’s card shows Transformers with trust-remote-code and vLLM, and the model can also be called through the NVIDIA Build API or Baseten. Its recommended maximum input is 1,664 by 2,048 pixels, so PDFs must be rendered to page images first. The card lists Linux with NVIDIA H100 or H200 GPUs as the tested setup.

License

Both models use the OpenMDW-1.1 license, a short permissive license for machine-learning models from the Linux Foundation and the PyTorch Foundation. It covers the model, its data, documentation and software together, and it puts no restrictions on outputs. The cards say the models are ready for commercial or non-commercial use.

Why it matters

Together they cover two halves of a document pipeline: Parse extracts text, and Embed searches pages as images. LightOnOCR-3 is the closest open parser in NVIDIA’s table, and Qwen3-VL-Embedding is the multimodal embedding model to compare with Embed. The leaderboard positions are dated and the speeds were measured on NVIDIA hardware, so test both on your own documents.

The OCR and document models and embedding models lists show where they fit, and how to choose an embedding model explains what to test.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Nemotron 3.5 Embed NVIDIA 1.55B – 8.38B 32K OpenMDW-1.1 (Permissive) 1.46 GB Laptop or phone
Nemotron 3.5 Parse NVIDIA 631M – OpenMDW-1.1 (Permissive) 0.435 GB Laptop or phone
LightOnOCR-3 LightOn 853M – 4.54B 256K Apache 2.0 (Permissive) 0.589 GB Laptop or phone
Qwen3-VL-Embedding Alibaba Qwen 2.13B – 8.14B 32K Apache 2.0 (Permissive) 4.9 GB Laptop or phone