nemotron-ocr-v2

by nvidia

28x faster multilingual OCR — production-grade throughput for RAG pipelines

2.9Kdl/month

N/Aparams

HuggingFace Use in Pipeline

Identifiers

Model ID

nvidia/nemotron-ocr-v2

Feature URI

mixpeek://image_extractor@v1/nvidia_nemotron_ocr_v2

Overview

Nemotron OCR v2 is NVIDIA's high-throughput OCR model designed for production RAG pipelines. At 34.7 pages per second on an A100, it processes documents 28x faster than PaddleOCR while supporting English, Chinese, Japanese, Korean, and Russian in a single architecture — no language detection step required.

The model uses a RegNetX backbone for visual feature extraction paired with a Transformer decoder for text generation. On Mixpeek, it powers bulk document ingestion where throughput is the bottleneck — scanning millions of pages into searchable text at speeds that keep up with real-time document feeds.

Architecture

RegNetX visual backbone with Transformer text decoder. Unified architecture handles 5 languages (en, zh, ja, ko, ru) without language detection. Optimized for batch inference with TensorRT acceleration.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mx = new Mixpeek({ apiKey: "API_KEY" });

await mx.collections.ingest({
  collection_id: "my-collection",
  source: { url: "https://example.com/multilingual-report.pdf" },
  feature_extractors: [{
    name: "ocr",
    version: "v1",
    params: {
      model_id: "nvidia/nemotron-ocr-v2"
    }
  }]
});