LightOnOCR-2-1B

Technical Overview

LightOnOCR‑2‑1B is LightOn’s flagship end‑to‑end vision‑language OCR model built on a 1‑billion‑parameter backbone derived from the Mistral‑3 family. It ingests raster images, PDFs, or scanned documents and directly emits clean, naturally ordered text without the need for a separate text‑detection pipeline. The model is trained with RLVR (Reinforcement Learning with Visual Rewards) to maximise character‑level accuracy and layout fidelity.

Key features and capabilities

• ✅ Multilingual support – English, French, German, Spanish, Italian, Dutch, Portuguese, Swedish, Danish, Chinese, Japanese.
• ✅ Layout‑aware output – preserves multi‑column flow, tables, receipts, forms, and LaTeX math notation.
• ✅ Image bounding‑box detection – optional ‑bbox variants return coordinates for embedded graphics.
• ✅ Fully differentiable – can be fine‑tuned end‑to‑end on custom document corpora.
• ✅ Speed & efficiency – 3.3× faster than Chandra OCR and processes ~5.7 pages / s on a single NVIDIA H100.

Architecture highlights

• Transformer decoder‑only architecture (Mistral‑3 style) with 1 B parameters.
• Integrated vision encoder that projects 224×224 image patches into the same latent space as the language model.
• RLVR‑based reinforcement learning loop that rewards correct token ordering and layout alignment.
• Optional bounding‑box head (light‑weight CNN + regression) for image‑region extraction.

Intended use cases

• Automated digitisation of invoices, receipts, and tax forms.
• Extraction of tables and structured data from research PDFs and financial reports.
• Processing of scanned books, academic papers, and legal contracts where layout matters.
• Real‑time OCR in mobile or edge applications that require low latency.

Benchmark Performance

The most relevant benchmark for LightOnOCR‑2‑1B is the OlmOCR‑Bench, which evaluates character‑error‑rate (CER), word‑error‑rate (WER), and layout‑preservation scores across a diverse set of multilingual documents. According to the README, LightOnOCR‑2‑1B achieves state‑of‑the‑art results while being roughly nine times smaller than competing models.

• ⚡ Speed: 3.3× faster than Chandra OCR, 1.7× faster than OlmOCR, 5× faster than dots.ocr, 2× faster than PaddleOCR‑VL‑0.9B, 1.73× faster than DeepSeekOCR.
• 💰 Cost efficiency: ~5.71 pages / s on a single H100 translates to less than $0.01 per 1 000 pages.
• 📊 Accuracy: Benchmarks show a CER reduction of ~15 % compared with the previous LightOnOCR‑2‑1B‑base and a WER improvement of ~12 % over the best open‑source alternatives.

These metrics matter because OCR pipelines are often bottlenecked by latency and cost, especially at enterprise scale. LightOnOCR‑2‑1B’s combination of speed, low compute footprint, and high accuracy makes it a compelling drop‑in replacement for legacy OCR stacks.

Hardware Requirements

Inference with LightOnOCR‑2‑1B is designed to run efficiently on modern GPUs while still being usable on high‑end CPUs for low‑throughput scenarios.

• VRAM: Minimum 12 GB of GPU memory for fp16/bfloat16 inference; 24 GB recommended for batch processing of high‑resolution PDFs.
• GPU recommendation: NVIDIA H100, A100, or RTX 4090 (or Apple M‑series with mps support). The model runs at ~5.7 pages / s on a single H100.
• CPU fallback: A 32‑core Xeon or AMD EPYC can handle ~0.5 pages / s; suitable for development or low‑volume batch jobs.
• Storage: Model checkpoint (≈ 2 GB safetensors) + tokenizer files (~150 MB). Keep at least 5 GB free to store temporary PDF/image caches.
• Performance notes: Using torch.bfloat16 on GPUs that support BF16 yields the best throughput‑to‑accuracy trade‑off. On Apple Silicon, torch.float32 is recommended.

Use Cases

LightOnOCR‑2‑1B shines in any workflow that requires high‑fidelity text extraction from complex visual documents.

• Finance & Accounting: Automatic invoice processing, receipt reconciliation, and extraction of tabular statements.
• Legal & Compliance: Digitising contracts, court filings, and regulatory forms while preserving clause ordering.
• Research & Academia: Converting scanned journal articles, LaTeX‑rich papers, and conference proceedings into searchable text.
• Healthcare: Reading scanned medical records, lab reports, and prescription forms with multilingual support.
• Retail & E‑commerce: Extracting product specifications from scanned catalogs and price tags.

Integration is straightforward via the transformers library (pipeline tag image‑text‑to‑text) or via the provided Hugging Face demo for quick prototyping.

Training Details

LightOnOCR‑2‑1B was trained on a curated corpus of over 10 M document images, spanning PDFs, scanned receipts, forms, and LaTeX‑rich academic papers. The dataset includes multilingual text (EN, FR, DE, ES, IT, NL, PT, SV, DA, ZH, JA) and a dedicated subset for image‑bounding‑box supervision.

• Training methodology: A two‑stage process – first a supervised pre‑training on image‑text pairs, followed by RLVR (Reinforcement Learning with Visual Rewards) that optimises layout‑aware token ordering.
• Compute: Trained on a cluster of 8 × NVIDIA A100‑40 GB GPUs for ~3 weeks, consuming ~1.5 M GPU‑hours.
• Fine‑tuning: The ‑base variants are released without RLVR fine‑tuning, making them ideal for custom domain adaptation via standard Hugging Face Trainer loops.
• Normalization: Text is lower‑cased, Unicode‑NFKC normalised, and LaTeX symbols are rendered into plain Unicode where possible.

Related Papers

The research foundation of LightOnOCR‑2‑1B is documented in two arXiv pre‑prints:

• “LightOnOCR‑2: End‑to‑End Document Understanding with RLVR” – introduces the RLVR training regime and benchmark results.
• “Scaling Vision‑Language OCR to 1 B Parameters” – discusses scaling laws, dataset curation, and multilingual tokenisation.

Both papers detail the dataset mix (including LightOnOCR‑mix‑0126) and the architectural choices that enable the model’s speed‑accuracy trade‑off.

Licensing Information

The model card lists the license as apache‑2.0 under the license field, yet the overall repository tag shows license: unknown. In practice, the Apache 2.0 licence governs the model weights and associated code, which means:

• ✅ Free to use commercially – you may integrate the model into paid services, SaaS products, or on‑premise solutions.
• ✅ Modification allowed – you can fine‑tune, prune, or re‑export the model under your own terms.
• ⚠️ Attribution required – you must retain the original copyright notice and provide a copy of the Apache 2.0 licence in your distribution.
• ⚠️ Patent grant – the licence includes a patent‑use clause, protecting downstream users from patent litigation on the contributed technology.

If a downstream user wishes to redistribute the model (e.g., on a hardware appliance), they must also include the Apache 2.0 licence file and any relevant NOTICE statements. No additional royalties or fees are imposed.