LightOnOCR-2-1B: a 1B-parameter open OCR model hits SOTA at under $0.01 per 1,000 pages
IgorCarron · x · 2026-10-09
France's LightOn released LightOnOCR-2-1B, an Apache 2.0 open OCR model that beats rivals 9x its size at document parsing.
- Performance: state-of-the-art on OlmOCR-Bench; inference 3.3x faster than Chandra, 5x faster than dots.ocr, 1.7x faster than OlmOCR
- Throughput & cost: 5.71 pages/sec on a single H100 (493k pages/day per card), under $0.01 per 1,000 pages
- Architecture: fully end-to-end — PDFs, scans, or photos in, correctly ordered structured text out, avoiding the ordering failures of multi-stage pipelines (layout segmentation → detection → recognition → stitching); handles multi-column layouts, spanning tables, forms, and outputs clean LaTeX for math
- Deployment: 11 languages, native vLLM/SGLang support, runs locally via Ollama/LM Studio — well suited for mass PDF ingestion for RAG or document-entry pipelines
Related event: LightOnOCR-2: tiny 1B open model beats rivals 9x its size(2 posts)→
More from Multimodal
- Epic ships official Unreal MCP, letting Claude Code and Cursor drive the Unreal Editor — Scobleizer · 2026-10-09
- ByteDance's Dreamina launches AI film label with up to 15M credits and $10K promotion per project — lmoroney · 2026-10-09
- Score Studio runs a "leaked" GTA 6 dirt race, catching every overtake and dust plume — markjeffrey · 2026-10-09
- Building Rome from a single image: new method reconstructs 3D scenes with geometry beyond the visible view — jonstephens85 · 2026-10-09
- LightOnOCR-3 Seamlessly Extracts Hard-to-Read German Text in Demo — IgorCarron · 2026-10-09
- Open-source local image library asks: how do you find an old generation by its settings? — shivam_dewan · 2026-10-09