New OCR leaderboard tests 14 open models on 2,165 historical pages; best hits 97.6% accuracy under $2 per 1k pages
IgorCarron · x · 2026-10-24
The FineBooks project, with Hugging Face, EleutherAI, and storytracer, released a leaderboard scoring 14 open OCR models on 2,165 pages of expert human transcriptions of historical books.
Key results:
- Best model reads historical print at 97.6% character accuracy for under $2 per 1,000 pages
- Chandra-OCR-2 ranks 3rd of 20, close behind dots.ocr
- LightOnOCR-3-4B ranks 5th and hallucinates far less text on blank pages than LightOnOCR-2
- Indra OCR runs on CPU and ties Tesseract
The goal is to identify models good enough to re-OCR public domain books and unlock a rich data source for AI and researchers.
More from Models
- Moonshot's Kimi slashes usage quotas, walks back after backlash — limits now roughly 6x higher — teortaxesTex · 2026-10-24
- Do AI benchmarks reflect real workflows? Reddit thread flags six methodology flaws — Lucky_Creme_5208 · 2026-10-24
- Screenshot of a model's candid meta-cognitive answer on what it means to understand a question — zalxi · 2026-10-24
- Leaked Claude UI screenshot shows unannounced Opus 5.5, Opus 5 and Fable 5.1 models — adonis_singh · 2026-10-24
- Developer on Claude Opus 5.5: "nothing else even comes close right now" — prasenx · 2026-10-24
- NYU/UT Austin paper: LLMs write convoluted text due to length-penalty optimization — menhguin · 2026-10-24