LightOnOCR-1B-1025-AIO-GGUF

LightOnOCR-1B-1025-AIO-GGUF

The LightOnOCR-1B-1025 from lightonai is a compact 1B-parameter end-to-end vision-language model optimized for Optical Character Recognition (OCR) and document understanding, combining a Pixtral-based Vision Transformer encoder with a lightweight Qwen3-based text decoder distilled from high-quality VLMs, achieving state-of-the-art accuracy in its class on Olmo-Bench (76.1% overall, with 81.4% on ArXiv papers, 71.6% on old scans, and strong math/table handling) while processing 5.71 pages/second on a single H100 GPU—5× faster than dots.ocr, 2× faster than PaddleOCR-VL-0.9B, and 1.73× faster than DeepSeekOCR at under $0.01 per 1,000 pages. Fully differentiable without external OCR pipelines, it excels at layout-aware extraction from high-resolution PDFs (rendered at ~1540px longest dimension), handling tables, receipts, forms, multi-column layouts, math notation, and multilingual text (Latin-dominant) under Apache 2.0 license, with compact 32k/16k vocab variants for European languages and vLLM/Transformers support for efficient inference and LoRA fine-tuning.

LightOnOCR-1B-1025 [GGUF]

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):