Nemotron 3.5 Parse

NVIDIA's 631M page parser that returns text, tables, chart data, reading order and boxes. ParseBench 75.20, OpenMDW-1.1.

In short. Nemotron 3.5 Parse looks at a picture of a document page and writes out its text, tables and chart data in reading order, with the position of each element. It is small, at 631 million parameters, and companies can use it commercially. The card lists Linux and NVIDIA H100 and H200 GPUs as its tested setup.

Summary

NVIDIA published Nemotron 3.5 Parse on Hugging Face on 17 October 2026 and described it on 21 October in a blog post alongside Nemotron 3.5 Embed and Rerank. It turns an image of a page into structured output: text in Markdown, tables, chart data, element classes such as headings and captions, reading order and bounding boxes, exported as Markdown, HTML or JSON. The model has 631 million parameters. It combines a C-RADIOv4-SO400M vision encoder with a 10-layer mBART decoder and includes a multi-token-prediction head for faster decoding, and an 89-million-parameter DFlash2 drafter for speculative decoding was added on 4 November. NVIDIA evaluated its OCR on 163 language and script entries.

You choose how much work each page gets. A single pass parses the whole page once, a double pass re-parses detected tables and charts at higher detail, and a triple pass also runs a separate 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 score is the highest in NVIDIA's comparison, but only 0.1 points above LightOnOCR-3-4B at 75.10. The other models' scores are their published figures, while the pages-per-second numbers are NVIDIA's own measurements, and in NVIDIA's table Parse processes 13.5 pages per second against 2.8 for LightOnOCR-3-4B.

The license is OpenMDW-1.1, with the bundled tokenizer under MIT, and the card says the model is ready for commercial or non-commercial use. The BF16 weights are about 2.7 GB. The card shows Transformers (with trust-remote-code) and vLLM, and it can also be called through the NVIDIA Build API or Baseten. The recommended maximum input is 1,664 by 2,048 pixels, so PDFs must be rendered to page images first.

Good for

  • Turning scanned or rendered PDFs, slides and forms into Markdown, HTML or JSON, with reading order and bounding boxes
  • High-volume parsing, since one, two or three passes trade accuracy for speed
  • Search pipelines that pair it with Nemotron 3.5 Embed and Rerank
  • Table-heavy documents, where NVIDIA's triple-pass table score is 84.67

Watch out for

  • Chart extraction is the weakest of NVIDIA's category scores, at 64.40 against 84.67 for tables
  • The output is a string with control tokens and normalized coordinates, so you need NVIDIA's postprocessing helpers
  • Input is page images only, with a recommended maximum of 1,664 by 2,048 pixels
  • The comparison scores for other models are published figures, and the speed numbers are NVIDIA's own
  • The card lists Linux and NVIDIA H100 and H200 GPUs as the tested setup

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Nemotron 3.5 Parse variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Nemotron 3.5 Parse BF16 631M 1.45 GB 0.726 GB 0.435 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs.

More about OpenMDW-1.1

See also

Frequently asked questions

What is Nemotron 3.5 Parse?
Nemotron 3.5 Parse looks at a picture of a document page and writes out its text, tables and chart data in reading order, with the position of each element. It is small, at 631 million parameters, and companies can use it commercially. The card lists Linux and NVIDIA H100 and H200 GPUs as its tested setup.
Can I use Nemotron 3.5 Parse commercially?
Yes. Nemotron 3.5 Parse is released under OpenMDW-1.1, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Nemotron 3.5 Parse open source?
Nemotron 3.5 Parse is open-weight, not open source in the strict sense. You can download the model files, but the OpenMDW-1.1 adds conditions that the Open Source Initiative definition does not allow.
Can I run Nemotron 3.5 Parse on my own computer?
Yes. The smallest variant, Nemotron 3.5 Parse, needs about 0.435 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Nemotron 3.5 Parse released?
Nemotron 3.5 Parse was released by NVIDIA on October 17, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
LightOnOCR-3 LightOn 853M – 4.54B 256K Apache 2.0 (Permissive) 0.589 GB Laptop or phone
PaddleOCR-VL Baidu 960M – Apache 2.0 (Permissive) 0.662 GB Laptop or phone
Chandra OCR 2 Datalab 5.3B 256K OpenRAIL-M (Datalab) (Conditional) 3.66 GB Laptop or phone
Nemotron 3.5 Embed NVIDIA 1.55B – 8.38B 32K OpenMDW-1.1 (Permissive) 1.46 GB Laptop or phone