Tiny 15.9M-parameter Kraken OCR model beats far larger VLMs

Kraken's PP-OCRv6, a specialized OCR model with only 15.9M parameters, ranked 4th in CER on a 2,165-page historical document benchmark, outperforming VLMs hundreds of times larger, and excelled at preserving long s, ligatures and other historical text features.

2026-09-03 ~ 2026-09-03 · 2 related posts