What Nemotron OCR v2 does
Nemotron OCR v2 is NVIDIA’s current OCR model for the NeMo Retriever ecosystem. OCR, or optical character recognition, converts text visible in an image into machine-readable text. Nemotron OCR v2 extends that task by identifying where text appears and modeling relationships between detected regions. This helps downstream systems preserve more of a document’s structure instead of receiving an unorganized string of characters.
The model is designed for images such as scanned documents, receipts, forms, business records, image-only PDF pages, charts, infographics, and natural-scene text. Its output can include recognized text, bounding boxes showing the location of each detected region, and confidence scores. Depending on the inference configuration and deployment path, results can be aggregated at word, sentence, or paragraph level.
In practical terms, the model is best understood as an image-to-structured-text component. It can prepare documents for search, retrieval-augmented generation (RAG), indexing, or agentic workflows, but it is not itself a general-purpose chatbot or document-reasoning system.
Where it fits in NVIDIA’s current lineup
Nemotron OCR v2 belongs to NVIDIA’s NeMo Retriever family, which provides components for extracting information from documents and other sources used in retrieval systems. The current OCR documentation identifies Nemotron OCR v2 as the supported OCR model, with English and multilingual variants. NVIDIA makes it available through NVIDIA NIM deployment options and as downloadable model weights for local use.
The two documented variants target different language and recognition requirements:
| Variant | Recognition capacity | Character set | Language coverage |
|---|---|---|---|
| English | Maximum sequence length of 32 | 855 characters | English |
| Multilingual | Maximum sequence length of 128 | 14,244 characters | English, Simplified Chinese, Traditional Chinese, Japanese, Korean, and Russian |
The multilingual variant is the broader choice for mixed-language collections and is identified as the default in the current NIM support documentation. The sequence lengths above describe the recognizer’s text capacity; they should not be interpreted as a conversational context window or a maximum document length for an entire workflow.
How the model works
Nemotron OCR v2 uses a hybrid detector-recognizer design. A RegNetX-8GF detector first locates text regions in the image. A pre-norm Transformer recognizer then transcribes the text found in those regions. A separate relational module models logical grouping, reading order, and layout relationships.
This division of responsibilities matters for documents where visual position carries meaning. For example, a receipt may contain a store heading, item descriptions, prices, and totals in distinct regions. A form may place labels beside fields, while a business document may contain multiple columns or a table. Detecting text alone does not explain those relationships; the model’s layout analysis is intended to provide additional structure for systems that reconstruct or index the document.
NVIDIA documents 53,831,335 parameters for the English variant and 83,853,044 parameters for the multilingual variant. These are model-size specifications, not a guarantee of accuracy or processing speed for a particular hardware configuration.
Inputs and outputs
The model accepts RGB images in PNG or JPEG form. Inputs can be represented as float32 or uint8 tensors and can be supplied as a single image with dimensions of 3 × H × W or as a batch with dimensions of B × 3 × H × W. The documented preprocessing automatically converts uint8 values and applies multi-scale resizing.
The direct output is structured OCR information rather than a newly generated image, audio file, or video. Typical returned information includes:
- Detected text regions and their bounding boxes
- Recognized text for each region
- Confidence scores
- Logical grouping and reading-order information, depending on configuration
Nemotron OCR v2 therefore supports image input and text-based structured output. It does not natively produce images, audio, video, embeddings, or speech. There is no supplied documentation for a conventional language-model context window or a maximum output-token limit. The published recognizer sequence lengths are the relevant documented recognition constraints.
Deployment options and runtime support
NVIDIA provides Nemotron OCR v2 through NIM and as downloadable model weights. The supported deployment documentation covers NVIDIA GPU systems based on Ampere, Blackwell, Hopper, and Lovelace architectures, with Linux identified as the preferred operating system. The model card identifies TensorRT and PyTorch as supported runtime engines, while the downloadable repository provides a PyTorch-based local inference path.
This makes the model a better fit for organizations that already operate NVIDIA GPU infrastructure than for users seeking a lightweight, hardware-independent OCR library. Official optimized deployment may require compatible NVIDIA hardware and the associated runtime environment. Before production use, teams should test image sizes, document layouts, language mixtures, and post-processing requirements on their own data.
Pricing and availability
No official public per-token, per-image, or recurring usage price for the exact Nemotron OCR v2 model was identified in the supplied sources. As a result, a reliable price comparison cannot be calculated from the available documentation. Costs may depend on the selected NVIDIA distribution, infrastructure, deployment arrangement, and GPU capacity rather than on a published model-level price.
The model is available through NVIDIA’s Build.NVIDIA.com model documentation, NIM-related deployment documentation, and the NVIDIA model repository on Hugging Face. Availability of a downloadable model does not by itself establish that hosted inference is free or that every deployment path has identical terms.
Main strengths and trade-offs
Nemotron OCR v2’s primary strength is specialization. Its architecture addresses both recognition and document layout, which is more useful for ingestion pipelines than plain text transcription alone. The multilingual variant also covers six language groupings, including Chinese scripts, Japanese, Korean, and Russian, making it suitable for multilingual collections that would exceed the scope of the English variant.
Its deployment options are another practical advantage for NVIDIA-based environments. TensorRT support can be relevant to optimized inference, while PyTorch weights provide a more accessible local development path for teams building their own processing pipeline.
The trade-off is that this specialization does not make the model a replacement for a general multimodal language model. Nemotron OCR v2 extracts and organizes visible text; downstream software remains responsible for document reconstruction, schema normalization, indexing, validation, and higher-level interpretation. A separate system may be needed to answer questions about the extracted content or perform complex reasoning over it.
The editorial assessment supplied with this model rates its speed and cost favorably relative to its specialized role, while assigning low reasoning and coding scores. These are editorial evaluations, not NVIDIA benchmark results or vendor-published ratings. They reflect that an OCR-focused model can be efficient for its intended task but is not designed for code generation or open-ended reasoning.
Best use cases
- Extracting text from scanned documents and image-only PDF pages
- Preparing receipts, forms, invoices, and business records for search or indexing
- Adding OCR preprocessing to retrieval-augmented generation systems
- Preserving text locations and layout relationships in tables, charts, and infographics
- Building multilingual image-text retrieval pipelines
- Supplying structured document content to downstream agents or business workflows
For a RAG pipeline, Nemotron OCR v2 can serve as the first processing stage: it detects and transcribes text, returns locations and confidence information, and passes normalized content to chunking, indexing, and retrieval components. An application can use confidence scores to flag uncertain regions for review instead of treating every recognized character as equally reliable.
When to choose Nemotron OCR v2
Choose Nemotron OCR v2 when the input is primarily visual document content and the output needs to retain more than plain text. It is particularly appropriate when multilingual recognition, bounding boxes, reading order, or layout-aware extraction matters, and when the deployment environment already includes supported NVIDIA GPUs.
The multilingual variant is the sensible starting point for collections containing Chinese, Japanese, Korean, or Russian alongside English. The English variant may be appropriate when the corpus is limited to English and the smaller recognizer is sufficient. The supplied research does not provide a direct accuracy comparison between the variants, so selection should be validated against representative documents.
Another option may be more appropriate when the main requirement is conversational question answering, code generation, broad visual reasoning, embeddings, or native image, audio, or video generation. Likewise, teams without compatible NVIDIA GPU infrastructure may prefer a different OCR service or library if minimizing deployment complexity is more important than using NVIDIA’s supported stack.
Limitations to check before production
Nemotron OCR v2 does not provide a documented general-purpose knowledge cutoff, because it is an OCR model rather than a conversational model trained around a stated knowledge boundary. The supplied information also does not verify tool use, function calling, streaming, fine-tuning, caching, batch APIs, or a distinct JSON mode. Applications should not assume those capabilities merely because the OCR results are structured.
Language coverage differs substantially between the English and multilingual variants, and the recognizer sequence capacities also differ. Deployment is tied to supported NVIDIA GPU environments for the documented optimized paths. Finally, OCR confidence is not the same as application-level correctness: documents should be tested for small text, unusual layouts, tables, mixed scripts, and image quality before automated extraction is trusted.