What changed
Daniel van Strien of Hugging Face compared eight OCR models that recognize text in images, testing them on 850 pages from old natural history books, the 1771 Encyclopaedia Britannica, and typed and printed archival documents. All models ran on the same GPU, an Nvidia RTX PRO 6000 rented for $2.75 per hour. The oldest model from April 2025 (RolmOCR) cost about 51 cents per 1,000 pages, while the newest from October 2026 (LightOnOCR-3-0.8B) came in at just under 10 cents. A million pages now runs between $65 and $130 in GPU rental, depending on the source material.
What this means in practice
Optimizing how the models ran saved about as much money as switching to newer models did, if not more. This included feeding pages more evenly, using smaller scans, updating software, and running a faster GPU. The newest model read book pages at least as accurately as the oldest one. However, some models hallucinated text on blank pages or repeated the same content over and over.
Limitations and availability
The supplied announcement does not specify the exact workflow or access route for these models. The cost savings are based on GPU rental rates and model performance on specific test materials. The models tested include RolmOCR and LightOnOCR-3-0.8B, with five models in between almost on par with each other. The accuracy of the newest model was at least as good as the oldest, though some models had issues with blank pages and repeated content.
