LightOnOCR-2-1B: 1B OCR model beats rivals 9x its size, 493K pages/day on one GPU
thisguyknowsai · x · 2026-09-18
French AI lab LightOn released LightOnOCR-2-1B (Apache 2.0), a 1B-parameter OCR model that turns PDFs, scans, or photos into clean, correctly ordered text with a single end-to-end model — no brittle multi-stage pipeline.
- State-of-the-art on OlmOCR-Bench while being 9x smaller than competing models.
- 5.71 pages/sec on a single H100 (493K pages/day, under one cent per 1,000 pages); 3.3x faster than Chandra, 5x faster than dots.ocr, 1.7x faster than OlmOCR.
- Handles tables, receipts, forms, multi-column layouts, and math notation with clean LaTeX output; supports 11 languages.
- Runs on vLLM, SGLang, Ollama, or LM Studio.
More from Models
- PrismML launches ternary Bonsai 2 27B: 5.9GB, 9x smaller, retains 98.2% of full-precision performance — airesearch12 · 2026-09-18
- Grok Voice Transcribe 2.0 hits 92.9% accuracy on phone calls at just $0.10/hour — XFreeze · 2026-09-18
- Meta's 'Muse Spark' spotted on Hugging Face, release appears imminent — eliebakouch · 2026-09-18
- Silent Model Revisions Break Everything — Versioning and Model Cards Are a Mess — xeophon · 2026-09-18
- Suspected Gemini 4 Pro generates animated peacock-on-a-bike SVG in Arena — CodeByPoonam · 2026-09-18
- Suspected Gemini 4 Pro builds a Three.js rocket from a single-line prompt — CodeByPoonam · 2026-09-18