Build a Fast Multilingual OCR with Nemotron-OCR-v2

NVIDIA’s April 17, 2026 release of the Nemotron-OCR-v2 model provides a production-ready system for high-speed document and scene text extraction. This 83.9 million parameter model replaces the English-only v1 release with support for six major languages and advanced structural analysis. You can use it to process up to 20 pages per second. This guide covers the architecture, deployment requirements, and how the model achieves its latency targets for high-volume document pipelines.

Hybrid Neural Architecture

Nemotron-OCR-v2 relies on a three-part neural network architecture to handle complex layouts. The system processes images end-to-end to capture both text and structural context.

The primary modules include:

• Detector: A convolutional-based model for precise text region localization.
• Recognizer: A Transformer-based component that transcribes detected regions utilizing a 14,000-character vocabulary.
• Relational Model: A specialized module for layout and document structure analysis.

The Relational Model handles advanced reading order analysis. This structural awareness ensures that text extracted from complex formats like multi-column PDFs and dense tables maintains its logical flow. Maintaining structural integrity is a strict requirement when building a RAG application or feeding context to multimodal agents.

Synthetic Training Data Pipeline

Model accuracy and generalization rely heavily on synthetic data generation. The total training dataset consists of approximately 12 million images.

NVIDIA utilized over 11 million rendered multilingual document pages. This synthetic dataset covers English, Japanese, Korean, Russian, Simplified Chinese, and Traditional Chinese. The generation pipeline included historical document crops and archaic characters. Simulated degradation effects were applied to ensure the model handles poor-quality scans in production environments.

The remaining 680,000 images consist of real-world data. This subset targets natural scene text, charts, infographics, and table images equipped with bilingual annotations.

Throughput and Benchmarking

Processing speed defines this release. Independent benchmarking reports a latency of approximately 48.7ms per page.

This translates to a throughput of up to 20 pages per second on optimized hardware. Traditional competitor models often average around 0.6 pages per second. High throughput reduces compute costs for enterprise tasks like invoice and contract processing.

The model tops private leaderboards for OCRBench v2. It was also rigorously evaluated on OmniDocBench, a document OCR benchmark designed specifically for testing English, Chinese, and mixed-language content.

Deployment and Configuration

The model requires specific hardware and software configurations. You must deploy on Linux (amd64) systems equipped with the CUDA toolkit 12.x.

The multilingual OCR model is available via the Hugging Face Hub under the nvidia/nemotron-ocr-v2 repository. It is also packaged as part of the NVIDIA NeMo Retriever collection. The model weights fall under the NVIDIA Open Model License Agreement. All associated post-processing scripts are licensed under Apache 2.0.

Enterprise deployments can utilize NVIDIA NIM (Inference Microservices). Managing AI inference through NIM provides standardized API endpoints and predictable scaling for your internal applications.

Evaluate the Hugging Face repository and download the weights to test the model against your own complex document layouts. Review the Apache-licensed post-processing scripts to understand how the Relational Model structures its final text output.