What is Nemotron OCR v2?
Nemotron OCR v2 is a text-recognition model from NVIDIA's NeMo Retriever line, published on Hugging Face in April 2026. Like PaddleOCR's classic pipeline it is not a vision-language model: a RegNetX-8GF convolutional detector finds text regions, a small Transformer recogniser reads them, and a relational module groups them into lines and paragraphs in reading order. It returns boxes, text and confidence scores.
It ships in two variants. The English model has 53.8 million parameters and works at word level; the multilingual model has 83.9 million and reads whole lines in English, Simplified and Traditional Chinese, Japanese, Korean and Russian from one 14,244-character set. Training used about 680,000 real images plus over 11 million synthetic pages, released as a CC BY 4.0 dataset.
The weights are under the NVIDIA Open Model License, which allows commercial use. It needs Linux and an NVIDIA GPU, and the package compiles a CUDA extension at install time.
Is Nemotron OCR v2 worth it?
Nemotron OCR v2 trades accuracy for throughput, and NVIDIA's own numbers show both sides. On one A100 the multilingual model runs at 34.7 pages a second, against 1.2 for PaddleOCR v5 server and 1.5 for OpenOCR, which NVIDIA calls over 28 times faster. On the same OmniDocBench crop test its normalised edit distance is 0.048 on English, 0.072 on Chinese and 0.142 on mixed text, where PaddleOCR v5 scores 0.027, 0.037 and 0.041; lower is better, so Paddle is clearly more accurate on real documents.
The multilingual gain is real but measured on NVIDIA's ground. On synthetic SynthDoG pages it beats PaddleOCR in every language, with error falling from 0.923 to 0.047 in Korean against v1. Those test pages come from a generator close to NVIDIA's own synthetic training pipeline, whose modified code is not yet public, so treat the margin as an upper bound.
It also does less per page than the VLM parsers it sits beside. There is no table structure, no LaTeX and no Markdown, only boxes, text and reading order, and rotated pages hurt: the multilingual model's error on text rotated 270 degrees is 0.332. The English variant is useless on Chinese, at 0.830.
The deployment story is NVIDIA's strength. It ships as a NIM container, a hosted trial endpoint and a Docker Compose example, and the licence permits commercial use without revenue caps. A retrieval team indexing millions of mostly English pages can push them through one GPU in hours rather than days; a team that needs tables or the best character accuracy should pair it with, or replace it by, PP-OCRv6 or a VLM parser.
Where it earns its place
A platform engineer building a RAG index over 5 million internal PDFs
Extract text and reading order fast enough to re-index weekly
Run the multilingual model in the NIM container or the Python pipeline on A100s. At NVIDIA's 34.7 pages a second, one GPU handles about 125,000 pages an hour, so 5 million pages takes about 40 GPU-hours.
A developer at an East Asian e-commerce marketplace
Read product photos and labels in Chinese, Japanese and Korean
Use the multilingual variant, which reads all five languages with one model and no language-detection step, and keep confidence scores to flag low-quality images; check rotated images, which score worst in NVIDIA's table.
Before you commit
- Setup
- Clone the Hugging Face repo with git-lfs, install PyTorch matching your CUDA toolkit, then pip install --no-build-isolation in Python 3.12; or run the Docker Compose example with the NVIDIA Container Toolkit.
- Integrations
- nemotron_ocr Python API with detector-only and skip-relational modes; build.nvidia.com hosted endpoint; NIM container on NGC as part of NeMo Retriever.
- Data and privacy
- NVIDIA Open Model License, commercial use allowed; self-hosted runs keep images local. The hosted trial is governed by the NVIDIA API Trial Terms of Service.
- Cost at scale
- 34.7 pages a second (multilingual) or 40.7 (English) on one A100 in NVIDIA's batched pipeline; detector-only mode uses about 37% less GPU memory. NIM production pricing is not published on the model page.
- Limits
- Linux amd64 with Ampere or newer NVIDIA GPUs only; Python 3.12 exactly; hosted inline images must be under 180,000 base64 characters; recogniser sequences cap at 128 characters.
What’s good
- 34.7 pages a second on one A100 for the multilingual model
- 84M parameters with reading-order grouping and confidence scores
- Commercial use allowed under the NVIDIA Open Model License
- Synthetic training set of over 11 million pages released under CC BY 4.0
What’s not
- Higher error than PaddleOCR v5 on OmniDocBench in NVIDIA's own table
- No tables, formulas or Markdown output
- Linux and NVIDIA GPU only; compiles a CUDA extension at install
- Multilingual gains measured mostly on synthetic SynthDoG pages
Best for
Retrieval and RAG teams on NVIDIA hardware who need to extract plain text from very large page volumes quickly, with a commercial-friendly licence.
Nemotron OCR v2 alternatives
Nemotron OCR v2: common questions
- Is Nemotron OCR v2 worth it?
- Nemotron OCR v2 is the fastest open OCR here by an order of magnitude, but on NVIDIA's own document benchmark it reads less accurately than the PaddleOCR v5 it outpaces. We rate it 3.8 out of 5.
- How much does Nemotron OCR v2 cost?
- Free under the NVIDIA Open Model License, commercial use allowed. Trial API on build.nvidia.com; NIM container pricing not published on the model page.
- What are the downsides of Nemotron OCR v2?
- Higher error than PaddleOCR v5 on OmniDocBench in NVIDIA's own table; no tables, formulas or Markdown output; linux and NVIDIA GPU only; compiles a CUDA extension at install; multilingual gains measured mostly on synthetic SynthDoG pages.
- Who is Nemotron OCR v2 best for?
- Retrieval and RAG teams on NVIDIA hardware who need to extract plain text from very large page volumes quickly, with a commercial-friendly licence.
- How much does Nemotron OCR v2 cost for a team or in production?
- 34.7 pages a second (multilingual) or 40.7 (English) on one A100 in NVIDIA's batched pipeline; detector-only mode uses about 37% less GPU memory. NIM production pricing is not published on the model page.
- Is Nemotron OCR v2 safe for company data?
- NVIDIA Open Model License, commercial use allowed; self-hosted runs keep images local. The hosted trial is governed by the NVIDIA API Trial Terms of Service.
- What are the best Nemotron OCR v2 alternatives?
- The closest alternatives we have reviewed are PP-OCRv6, PaddleOCR-VL-1.6, Surya OCR 2.5 and DeepSeek-OCR 2.
Answers checked Nov 2, 2026.