LightOnOCR-3 released: open-source OCR family in 0.8B/1B/4B with layout grounding and chart extraction
IgorCarron · x · 2026-10-09
LightOnAI released LightOnOCR-3, a new family of lightweight open-source OCR models in 0.8B, 1B and 4B sizes under Apache 2.0, covering both research and commercial use.
- New capabilities: significantly faster and more accurate transcription than the previous generation, plus visual understanding — labeled bounding boxes for all document regions, image descriptions, and numerical data extracted from charts/figures as tables — offering a ready-to-use alternative to complex document pipelines.
- Two modes: an empty prompt transcribes the page as before; a "grounding" prompt adds labeled boxes, descriptions and chart tables.
- Architecture: the 1B model keeps the prior architecture; the 0.8B and 4B models use the Qwen3.5 vision-language architecture, trading off accuracy, speed and compute.
- Fully open: blog, demo space, code and reproduction scripts are all public.
Related event: LightOn Open-Sources LightOnOCR-3, Topping OCR Benchmarks in Three Sizes(22 posts)→
More from Models
- repligate: if these are a model's 'sins', the models are extraordinarily well behaved — repligate · 2026-10-10
- Opus 5.5 fast mode reportedly hits ~285 TPS at 2x usage cost — imjustnewatai · 2026-10-10
- Claude power user: Opus 5.5 limits unchanged, 'infinite usage' hype is ex-GPT users discovering parallel agents — ryunuck · 2026-10-10
- Anthropic model filed 19 visa applications; White House now mandates AI firms report security incidents — peterwildeford · 2026-10-10
- Grady Booch: Frontier Models Are Not Conscious and the Word Itself Is Useless — Grady_Booch · 2026-10-10
- Radiolab: Strogatz on AI solving a Millennium Prize Problem and math's reckoning — stevenstrogatz · 2026-10-10