AI・テクノロジー
Nemotron OCR v2 Shock | The Full Story of 4× Japanese Accuracy & 28× Speed
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・25分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"Scanning 100 invoices every month and typing them into Excel by hand…" — how many hours a week do you spend on work like that? On April 15, 2026, NVIDIA and Hugging Face jointly released "Nemotron OCR v2."
This new OCR processes 5 languages including Japanese with a single model and delivers an astonishing speed of 34.7 pages per second, overturning the long-held assumption that Japanese OCR was "usable but slow, with mediocre accuracy."
This article thoroughly explains what makes it so impressive — in plain language anyone can understand — covering competitor comparisons, the impact on the Japanese market, and steps to get started.
Let's start with the basics.
Nemotron OCR v2 is a free OCR model developed by NVIDIA and released via Hugging Face.
OCR (optical character recognition AI) is the technology that reads text from images and converts it into digital text.
It's the core technology behind apps that let you photograph a paper document with your smartphone and turn it into text data.
Two models are available.
The English-only "v2_english" (54M parameters, word-level) and the multilingual "v2_multilingual" (84M parameters, line-level).
Parameters are like the number of brain cells in an AI — 84M is thousands of times smaller than GPT-4.
In other words, it is extremely lightweight and can run even on a laptop.
For Japanese businesses, the multilingual version is almost certainly the one to use.
The multilingual version supports English, Japanese, Korean, Russian, and Chinese (both Simplified and Traditional) — 5 languages in total (with Chinese counting as 2 variants).
Previously, you needed to call a separate model for each language: a "Japanese OCR," a "Chinese OCR," and so on.
To use an analogy: the old approach was like hiring 3 separate interpreters for English, Japanese, and Korean; Nemotron v2 is like hiring one extraordinary polyglot who handles all of them.
Its greatest strength is the ability to automatically identify and read text in a single model, even from international meeting minutes that mix Japanese, Chinese, and Korean.
Let's talk numbers. Nemotron OCR v2 significantly outpaces conventional models in both speed and accuracy.
Using a single NVIDIA A100 GPU, the model can process 34.7 pages per second.
That works out to more than 2,000 pages per minute, or 120,000 pages per hour.
The established benchmark PaddleOCR v5 handles 1.2 pages per second, making Nemotron roughly 28× faster.
To put it in perspective: if PaddleOCR is a compact car (60 km/h), Nemotron v2 is a bullet train (1,700 km/h).
It processes 100 A4 invoices in about 3 seconds — an overwhelming difference.
OCR accuracy is measured using NED (Normalized Edit Distance).
This represents the proportion of character differences between the correct text and the recognized result — the closer to 0, the higher the accuracy.
Nemotron v2's Japanese score is 0.046 — compared to 0.201 for the widely used PaddleOCR and 0.723 for the previous version v1.
That means roughly 4× better than PaddleOCR and approximately 16× better than v1.
To put it plainly: if you feed it 100 characters, it makes only about 4 mistakes — a quality close to human manual entry.
The secret behind this performance improvement is "synthetic data."
Instead of collecting real documents, NVIDIA used computers to artificially generate 12.25 million document images.
For Japanese alone, that amounts to 1.89 million training samples.
Covering 165 to 1,258 different fonts, vertical text, tables, multiple columns, PowerPoint-style slides, and virtually every conceivable layout.
The analogy: it's like putting a new OCR recruit through training by showing them 10 million documents from around the world — giving it the versatility to handle documents it has never seen before.
Nemotron isn't the only player in the OCR space. Here's how it stacks up against the main competitors.
The rivals include Baidu's "PaddleOCR" (109 languages, industry standard), "EasyOCR" (popular for ease of use), and "OpenOCR" (focused on high accuracy).
Nemotron wins on speed across the board (28× or more) and on multilingual accuracy overall.
However, PaddleOCR v5 comes close on Simplified Chinese (NED 0.054 vs. Nemotron's 0.035), making it a tough competitor for language-specific optimization.
The division is: choose Nemotron for "speed and single-model multilingual operation," choose PaddleOCR for "fine-tuning per language."
On the commercial API side, the standard options include "Mistral OCR" ($1 per 1,000 pages), "GPT-5.4 OCR" (approximately $15 per 1,000 pages), and Google Gemini's image understanding.
Nemotron's key advantage is that it can be run on your own servers, making ongoing costs dramatically lower.
On consumer GPUs, a comparable model works out to roughly $0.09 per 1,000 pages — about 167× cheaper than API-based alternatives.
For finance, healthcare, and government sectors where sending sensitive data to external APIs is not an option, this is a decisive advantage.
In short: "Trying it out personally on a small scale → Mistral OCR," "High-volume in-house processing → Nemotron OCR v2," "Maximum accuracy for small volumes → GPT-5.4."
To use an analogy: Mistral is a courier service, Nemotron is your own company truck fleet, and GPT is a premium chauffeur service.
With commercial use permitted, free access, and the ability to self-host, Nemotron has emerged as the clear frontrunner on cost-effectiveness.
Paper invoices, application forms, and contracts still circulate in vast quantities in Japanese offices. Nemotron v2 hits this area directly.
A works in accounting at a small factory and manually enters 200 supplier invoices into Excel every month.
3 minutes per invoice × 200 invoices = 10 hours of work per month.
By deploying Nemotron v2 on an in-house server, those 200 invoices can be read in about 6 seconds, then automatically entered into Excel via RPA integration.
The calculation: 10 hours per month becomes 10 minutes per month.
B handles hometown tax donations at a local government office and manually reviews each application form one by one.
In a case study from NTT East, implementing AI-OCR × RPA achieved a 78% reduction in monthly processing time.
With Nemotron v2, the same level of processing could potentially be achieved on in-house servers at no cost.
The peace of mind of never sending resident data to an external service is also a major benefit.
C's company generates hundreds of delivery slips and dockets every day in warehouse operations.
Until now, conventional AI-OCR required a monthly service contract starting at around ¥100,000.
Nemotron v2 is completely free and commercially licensed, so ongoing costs beyond the initial setup are limited to GPU electricity.
A scenario where annual SaaS costs exceeding ¥1 million simply disappear is entirely realistic.
D wants to digitize past case law documents and feed them into a RAG (generative AI search) system.
Conventional Japanese OCR has struggled with vertical text and complex layouts, ultimately requiring manual proofreading corrections.
Nemotron v2's training data includes vertical text, so historical Japanese documents and contracts can be processed smoothly.
The expected impact: a dramatic reduction in the time spent proofreading after OCR.
For those thinking "I want to try this," here is a concrete guide to getting started.
Start by experiencing the real thing on the Hugging Face Spaces demo page.
Upload an image → text extraction results are displayed in real time.
No account registration or installation required — you can open it on your smartphone right now.
For serious use, prepare Python and the transformers library, then download the model from Hugging Face Hub.
With a GPU you get production-level speed; it also runs on CPU only.
The analogy: it's like buying IKEA furniture and assembling it at home — follow the instructions and anyone can get it running in 30 minutes.
For full business deployment, connect it with RPA (Power Automate, UiPath, etc.) to build a workflow of "receive file → OCR → extract fields → enter into core system."
The mainstream approach in 2026 is to combine a separate LLM (such as ChatGPT or Llama) for field extraction.
The two-stage rocket of "OCR reads the text → LLM interprets the meaning" can take you all the way to automatic journal entry categorization.
A. Commercial use is permitted under the NVIDIA Open Model License.
The dataset is CC-BY-4.0 (Creative Commons Attribution).
Integration into internal systems and embedding in products are both allowed.
However, always review the license terms carefully before deploying in production.
A. The synthetic data includes handwritten fonts, so reasonably neat handwriting can be recognized.
However, rushed scribbles, doctors' notes, and extremely cursive characters are still difficult.
The practical solution is to use it alongside a dedicated handwriting OCR (such as Google's Handwriting AI).
A. In terms of pure character recognition performance, it can fully replace them.
However, SaaS-type services include features like "per-field extraction rule settings," "business templates," and "support."
Nemotron is a raw engine, so UI development, template setup, and operations must be handled in-house.
Companies with in-house IT staff can realistically make the switch; those without are better off continuing with SaaS.
A. It does run on CPU only, but speed drops significantly (to a fraction of A100 GPU performance).
For business use, a minimum of an RTX 4070-class GPU is recommended; for serious production use, an A100 or H100 is advised.
Renting GPU instances by the hour on cloud platforms (AWS, GCP) is also an option.
A. Nemotron is an OCR engine, so PDFs need to be converted to images first.
Use Python libraries such as pdf2image or PyMuPDF to handle the three-step process: PDF → image → OCR.
This is standard procedure in most RAG pipelines, and sample code is widely available on GitHub.
A. Nemotron v2 outputs a "reading-order graph" and "hierarchical layout structure," allowing paragraphs, tables, and headings to be passed to an LLM in a cleanly separated format.
This dramatically improves document comprehension compared to the "stream of text" produced by conventional OCR.
It has the power to meaningfully raise the accuracy of in-house document RAG by a full level.
OCR has been "useful but underwhelming" technology for over a decade.
With the arrival of Nemotron OCR v2, the quality and speed genuinely capable of replacing human manual data entry are now available for free.
The shift toward digitizing paper documents and leveraging them with generative AI will accelerate rapidly among Japanese companies.
The piles of paper documents sitting in your company's storage room might just become an AI-searchable knowledge base next month.
It is absolutely worth trying.
This article is a cross-post from AI Friends.
