Newshot

AI Models Slash the Cost of Digitizing Historical Texts: Recognizing 1,000 Pages Now Costs Under $0.1

AI summary

With free, open AI models, the cost of extracting text from scanned books and archives has fallen sharply. Daniel van Strien of Hugging Face benchmarked eight image-to-text AI models on 850 pages, including old natural history books, the 1771 edition of the Encyclopaedia Britannica, and typed and printed historical archive documents, all run on the same 英伟达 (Nvidia) RTX PRO 6000 GPU rented via Hugging Face for $2.75 per hour. The earliest model tested, RolmOCR, released in April 2025, cost about $0.51 per 1,000 pages, while the newest, LightOnOCR-3-0.8B, released in October 2026, brought the cost to under $0.1; at that rate, recognizing 1 million pages would require roughly $65 to $130 in GPU rental, depending on the source material. Van Strien noted that optimizing how models are run can save as much as, or more than, switching to a newer model—through more balanced page-processing task allocation, smaller scan images, updated software, and faster GPUs—and that the newest models are at least as accurate as the earliest on book pages, though some models notably hallucinate text on blank pages or repeat the same output.

Translation(model translation, not the source's own text)

AI Models Slash the Cost of Digitizing Historical Texts: Recognizing 1,000 Pages Now Costs Under $0.1

IT之家 November 19 news: With the help of free and open AI models, the cost of extracting text from scanned books and archives has dropped sharply.

Daniel van Strien of Hugging Face ran a comparative test of 8 AI models capable of recognizing text in images. The test material totaled 850 pages, including old natural history books, a 1771 edition of the Encyclopædia Britannica, and typed and printed historical archive documents.

All models ran on the same GPU, namely the 英伟达 (Nvidia) RTX PRO 6000. Van Strien rented this graphics card through Hugging Face at a cost of $2.75 per hour (IT之家 note: about 18.4 yuan at current exchange rates).

The test results showed that, with identical image sizes, RolmOCR, the earliest model, released in April 2025, cost about $0.51 (about 3.4 yuan at current exchange rates) per 1,000 pages recognized, while the latest model, LightOnOCR-3-0.8B, released in October 2026, has brought the cost down to under $0.1 (about 0.67 yuan at current exchange rates). At this level, the GPU rental cost for recognizing 1 million pages is about $65 to $130 (about 434.3 to 868.5 yuan at current exchange rates), depending on the type of source material.

Van Strien noted that the savings from optimizing how the models run are comparable to, and possibly even greater than, the savings from switching to a newer-generation model. Specific optimization measures include distributing page-processing tasks more evenly, using smaller scanned images, updating software, and switching to faster GPUs.

In terms of recognition accuracy, the latest models perform at least as well as the earliest model when processing book pages. However, some models also have obvious problems, such as fabricating text on blank pages or repeatedly outputting the same content.

Cost comparison for scanning and recognizing 1,000 pages: the earliest model required about $0.51 (about 3.4 yuan at current exchange rates), the latest model has dropped to under $0.1 (about 0.67 yuan at current exchange rates), and the five models in between are basically at a similar level.

Related Articles

  • Online design and publishing tool Canva launches an "AI switch" that can turn off most AI tools
  • Huawei Pura X View wide-screen bar phone receives HarmonyOS 7.0.0.115 upgrade, with the desktop AI shortcut bar arriving as promised
  • 马斯克: SpaceX currently adds about 0.5GW of compute capacity per month, and will accelerate to 1GW per month by the end of next year
  • IMF Managing Director warns: excessive AI investment enthusiasm could lead to further adjustments in financial markets
  • Who is responsible when AI agents transfer money or make purchases without authorization? Multiple banks warn users may have to bear the losses themselves
  • Patriot launches AI Matrix series DDR5 memory and M.2 SSDs, designed specifically for on-device AI

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗