France's LightOn ships open-source LightOnOCR-3: 0.8B/4B models top OlmOCR-Bench and ParseBench
IgorCarron · x · 2026-10-09
French AI company LightOn launched LightOnOCR-3, a family of end-to-end document-understanding models (0.8B and 4B), succeeding a line with 4.5M+ Hugging Face downloads:
- Capabilities: one model call handles text/handwriting recognition, layout grounding with labels and bounding boxes, image description, and chart/scientific-figure data extraction; a prompt switch toggles between plain transcription and rich grounded output
- Benchmarks: first on FRBench-pdf2md (strong in handwriting), second on OlmOCR-Bench ahead of Chandra-OCR-2, Mistral OCR 4.1 and dots.mocr, and first among open-weight models on ParseBench
- Design: outputs are logically grouped, spatially grounded blocks that feed directly into chunking, retrieval and RAG pipelines
Model weights are available on Hugging Face.
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