Small Models Read Old Books Better Than Big Ones, and It Costs Two Dollars per Thousand Pages

Here is a problem you would not guess was holding AI back: the text of old books is bad. Libraries scanned millions of public domain volumes years ago and ran them through optical character recognition, the software that turns a picture of a page into letters. The results are full of errors, and models trained on that text learn badly. One project measured it: a language model trained on old machine read text learned at only 30 percent the efficiency of one trained on the same books transcribed by humans.

FineBooks, a joint effort by Hugging Face and EleutherAI, set out to check whether today’s open models can fix that. The team ran 14 openly available text recognition models over 2,165 pages of historical books and published the results as a public leaderboard. The best model, dots.mocr, reads 97.6 percent of characters correctly at 1.94 dollars per thousand pages. All 14 run on local hardware, no API key required.

The surprise is which models won. dots.mocr has 3 billion parameters, the internal numbers a model learns during training and a rough proxy for size. Qwen3.5-9B is more than three times larger and scores lower, at 94.9 percent. Second place goes to OvisOCR2 at 0.9 billion parameters, 96.9 percent accuracy, and 46 cents per thousand pages. For reading old books, size and quality simply do not track together.

The team is careful about what the numbers cover. The test uses only Antiqua typefaces in English, French, German and Latin, on single column pages. No Fraktur, no non Latin scripts, no handwriting. The ground truth comes from six volumes transcribed by experts between 2011 and 2012 at about one error per 2,000 characters. And the verdict splits by purpose: good enough for AI training data, not good enough for scholarship, because the models quietly modernise archaic characters such as the long s rather than misreading them. Libraries have a separate headache: their systems expect a format with word level coordinates, and these models output plain Markdown.

What this means for you: if you have ever scanned a document and fought with the garbled text that came out, the practical news is that free models now do this job well and run on your own machine. The wider point is nicer. FineBooks plans to reprocess about 200,000 public domain documents from the Biodiversity Heritage Library, which holds over 64 million pages of natural history, and release the clean text openly. That is a quiet, unglamorous improvement to the raw material every open model is built from, and the kind of work that rarely gets a press release.

Sources

• Hugging Face: the historical books OCR leaderboard
• The FineBooks leaderboard space
• The Decoder: old OCR text cripples language model training
• Biodiversity Heritage Library