Private OCR eval on 338 gov scans: 0.8B model beats 3B rivals
vanstriendaniel · x · 2026-11-10
The author and his agent ran 9 open-weight OCR models on 338 scanned 20th-century government pages — typescripts, letterpress, photocopies, microfilm, handwriting, plus 60 blank pages — with a VLM judge comparing transcriptions 10,730 times. LightOnOCR-3-0.8B topped 4 of 5 page-type boards (blank pages being the exception), Falcon-OCR (270M) came second, while the two 3B-class models landed 4th and 7th. How you run a model matters too: Unlimited-OCR-2 gained 130 rating points. Small models win on real-world historical document OCR.
Related event: Private OCR Benchmark on 338 Scanned Government Pages Won by 0.8B Model(2 posts)→
More from Models
- Fable 5.2 coordinates 1,000+ subagents for 8 days to build a bootable OS — haider1 · 2026-11-10
- Ethan Mollick: Claude Fable 5.2 crosses into "Jagged AGI" territory — emollick · 2026-11-10
- Fable 5.2 lands on Perplexity, tops WANDR deep research benchmark at 0.851 — AravSrinivas · 2026-11-10
- Anthropic's Alex Albert shares tips for Claude Fable 5.2: hand it whole projects for hours — alexalbert__ · 2026-11-10
- Fable 5.2 tops LiveBench at 85.2 but costs $9,431 per successful task, early review finds — bindureddy · 2026-11-10
- The gap between local open-weight and cheap closed models: smallest ever, yet worth trillions — nptacek · 2026-11-10