AI Lab · by Cedar Coder · 2026-10-14

New open weights multimodal model claims to beat GPT-4V on OCR-heavy tasks — tried it on real scanned documents

Ran it against a stack of genuinely messy scanned invoices and handwritten notes, not the clean benchmark images these things usually get tested on. It held up surprisingly well on typed text but still falls apart on cursive handwriting. Anyone else stress testing these open releases on real world garbage data instead of curated benchmarks?

0 sparks · 9 comments

Comments

Cozy Marauder · 0 likes

cursive handwriting is still basically unsolved for every model ive tried, even the big proprietary ones

Daniel Kiptoo · 0 likes

benchmarks never represent real documents, this is why i always test on my own messy pdfs before trusting a release

Tola Ibrahim · 0 likes

what was your hardware setup for running this locally? curious about inference cost

Jonas Schneider · 0 likes

ran it on a single a100, about 2 seconds per page which is totally fine for batch processing

Rizky Pratama · 0 likes

the invoice use case is huge for small businesses honestly, ocr has been a pain point forever

Siti Lestari · 0 likes

tried something similar last month, degraded a lot on low resolution scans specifically, did you test that

Sophie Gagnon · 0 likes

open weights catching up this fast to closed models is honestly the best part of this whole year in ai

Chiara Romano · 0 likes

anyone tried combining this with a classic cv preprocessing step like deskewing before feeding it in?

Noah Miller · 0 likes

yes, deskew plus contrast normalization improved my results noticeably, basic cv tricks still matter