LightOnOCR-2-1B

by lightonai

State-of-the-art 1B-parameter end-to-end multilingual OCR with bounding box localization

730Kdl/month

1Bparams

HuggingFace Use in Pipeline

Identifiers

Model ID

lightonai/LightOnOCR-2-1B

Feature URI

mixpeek://image_extractor@v1/lighton_ocr2_1b_v1

Overview

LightOnOCR-2-1B is a 1B-parameter vision-language model that sets the top score on OlmOCR-Bench (83.2) while being compact enough for efficient deployment. Built on a native-resolution ViT initialized from Mistral-Small-3.1, it handles page images up to 1540px on the longest edge with particularly strong performance on ArXiv papers, scanned documents with math, and complex tables.

On Mixpeek, LightOnOCR-2-1B extracts text from documents, scanned pages, and images with high accuracy, powering full-text search across document collections. An image-localization variant adds bounding box predictions without degrading OCR quality.

Architecture

Three-component VLM: native-resolution Vision Transformer (initialized from Mistral-Small-3.1) as encoder, multimodal projector, and language model decoder. Accepts page images up to 1540px longest edge. Optional bounding box localization via coordinate tokens introduced during pretraining and refined with RLVR.

Mixpeek SDK Integration

from mixpeek import Mixpeek
mx = Mixpeek(api_key="YOUR_KEY")

mx.ingest(
    collection_id="scanned-documents",
    source="s3://scans/",
    extractors=[{
        "type": "ocr",
        "model": "lightonai/LightOnOCR-2-1B",
        "output_feature": "extracted_text"
    }]
)