↻
Brett Bobley reposted
Daniel van Strien
@danielvanstrien.bsky.social
Small open OCR models are getting better at reading hard historic newspapers. Surya OCR 2.5 (689M) handled 1760s Fraktur, polytonic Greek and pre-reform Russian on my test pages, layout included. Big broadsheets need full-res input: error rate 25% → under 10% huggingface.co/da…
AI Weekly's analysis
→
- Surya OCR 2.5 is a 689M-parameter vision-language model handling OCR, layout analysis, table recognition, and reading order in one pass.
- It scores 85.9 on OmniParseBench and 83.7% on olmOCR-bench, up from 83.3 and 82.8% for Surya 2.
- On an H100 at 96 DPI throughput reaches 6.19 pages per second versus 2.28 for Surya 2, across 91 supported languages.
Read full analysis →
View on Bluesky →