RIP Commercial OCR. An Open-Source Model Just Topped Every Benchmark. | by Sumit Pandey | Apr, 2026 | Towards Deep Learning
A new open-source OCR model outperformed all major commercial tools on standard text and handwriting tests. It accurately transcribed a 1913 handwritten letter by Ramanujan, preserving layout, math notation, and faint ink details.
- New open-source OCR model (Vision Transformers + glyph-aware training) beat commercial tools like ABBYY and Google Cloud Vision on IFOCR, MLT, and ICDAR benchmarks, with word-error-rates below 1.5%—30-50% better than competitors
- Successfully transcribed a faded 1913 handwritten Ramanujan letter, correctly capturing layout, math notation, and faint ink strokes
- Trained on millions of scanned pages including medieval manuscripts and dense handwritten notes
- Released under Apache license with CPU/GPU deployment scripts, enabling free text extraction from receipts, legal documents, and historical archives without per-page fees
An open‐source OCR model just blew past every commercial benchmark, turning heads in fields from document scanning to historical research. A few months back, someone ran one of Srinivasa Ramanujan’s 1913 letters—faded ink, cramped math, century-old script—through this new system. It nailed the layout, captured the equations, even picked up the faint strokes that most proprietary tools miss.
Behind the scenes, the team built on a mix of Vision Transformers and custom glyph-aware training. They fine-tuned on millions of scanned pages, including medieval manuscripts and densely handwritten notes. On public tests—IFOCR, MLT, ICDAR—they posted word‐error‐rates below 1.5%, shaving off 30–50% compared to giants like ABBYY and Google Cloud Vision.
They’re shipping the weights under an Apache license and providing scripts for CPU and GPU deployment. That means you can extract searchable text from receipts, legal filings or historical archives without paying per page. Every benchmark they’ve thrown at it—printed, handwritten, low-light, multi‐language—ended up with this model on top. It’s a clear signal that open source is closing the gap on proprietary OCR in both accuracy and flexibility.