Open-Weight LLM · Private & Custom AI
LightOnOCR-1B-1025
End-to-end document OCR & parsing for ops teams: extract structured text from PDFs, receipts, forms, and tables at scale—self-hosted, no external pipeline needed.
LightOnOCR-1B is a 1.1B-parameter vision-language model purpose-built for optical character recognition and document understanding. It trades general reasoning for specialized speed (5–2× faster than competing OCR VLMs) and cost efficiency (~$0.01 per 1,000 pages). For ops, this means automating invoice processing, document ingestion, and knowledge extraction without reliance on third-party OCR vendors.
Model facts
Private deployment
Run LightOnOCR-1B-1025 in your own environment
Run entirely on your infrastructure using vLLM (v0.11.1+) or transformers. Single H100 processes ~5.71 pages/sec (~493k pages/day). Model weights are Apache 2.0 licensed; no external APIs required—all text extraction stays in your network. BF16 inference uses ~2.3 GB VRAM (estimated); pruned variants (32k, 16k vocab) fit tighter resource envelopes. Deploy as a containerized service behind your VPC.
Operational AI use cases
Invoice & Receipt Automation
Feed scanned/PDF receipts and invoices to extract line items, amounts, vendor names, and dates. Pipe structured output to accounting systems or expense databases. Handles multi-column layouts and tables natively—no manual OCR post-processing.
Document Triage & Knowledge Ingestion
Ingest legacy PDFs (contracts, policies, scientific papers, RFPs) and extract text for indexing into a vector DB or RAG system. LightOnOCR preserves layout awareness, so you retain section structure for downstream agents or retrieval pipelines.
Compliance & Forms Processing
Automate form reading (tax docs, applications, compliance questionnaires). Extract fields from handwritten and printed forms. Route parsed data to approval workflows or regulatory filing systems without manual data entry.
Custom AI
As a base for custom AI
Use as the document-understanding foundation for a custom AI application: train domain-specific LoRA adapters on your internal PDFs (financial statements, technical specs, vendor contracts) to improve accuracy for your vertical. Fully differentiable—suitable for continuous learning pipelines and fine-tuning with your proprietary document corpus.
In the operating system
Where it fits
Sits at the *knowledge intake layer* of an AI OS: converts unstructured documents into structured text feeding RAG, vector stores, and agent memory. Pairs with a lightweight LLM for post-extraction reasoning (e.g., Mistral or Qwen variants). Not a reasoning engine itself—a specialized *perception* layer for document-heavy workflows.
Data control & security
All OCR inference runs in your VPC; no document content sent to external OCR APIs. Architectural benefit: sensitive PDFs (contracts, health records, financial data) never leave your network. Model weights are open and auditable. *Note: self-hosting is an architecture choice; the model itself does not guarantee compliance or encryption—apply your own data governance and network security.*
Hardware footprint
BF16 (full precision): ~2.3 GB VRAM (estimate, excluding image cache). 32k/16k vocab variants fit ~1.5–1.8 GB. Inference latency ~500–800ms per page on H100. CPU fallback possible but very slow; GPU recommended. Batch inference scales linearly with vLLM's parallel processing.
Integration
vLLM server exposes OpenAI-compatible /v1/chat/completions endpoint—plug into existing API clients. Transformers pipeline works with standard PyTorch workflows. Accept PDFs or pre-rendered images (PNG/JPEG); render PDFs to 1540px longest dimension before inference. Batch processing supported. Output is raw text; post-process with regex or an LLM to extract structured fields (JSON, CSV). Compatible with LangChain, LlamaIndex, and custom Python pipelines.
When it's not the right fit
- —Handwriting-heavy documents: trained primarily on printed text; heavy cursive may degrade.
- —Real-time, sub-100ms latency: OCR is compute-bound; expect 500ms+ per page even on H100.
- —Non-Latin scripts (Arabic, CJK): model optimized for Latin alphabet; limited coverage for other writing systems.
- —Complex mathematical notation at scale: handles math but not tuned for dense equations like specialized math-OCR models.
Alternatives to consider
PaddleOCR-VL-0.9B
Lighter, faster for simple receipts/forms, but less accurate on complex layouts and no native RAG integration.
DeepSeekOCR
Stronger on handwriting and complex layouts, but 1.7× slower and larger VRAM footprint; less tuned for European languages.
LayoutLM (Microsoft)
Traditional layout-aware model; requires external OCR engine (Tesseract) + separate layout classifier; lower accuracy, but mature production libraries.
Related open models
FAQ
Can I run this entirely on-premises without internet?
Yes. Download model weights once from HuggingFace, then run vLLM or transformers in an isolated VPC. No callbacks to external APIs; all inference is local.
What are the commercial use rights?
Apache 2.0 license permits commercial use, derivative works, and distribution without royalty. Review the full license; no usage restrictions mentioned on the model card.
How do I fine-tune it for my domain (e.g., medical invoices)?
Model is fully differentiable. Use LoRA fine-tuning (see linked Colab notebook) with your annotated PDFs. Transformers integration for training is 'coming soon'; currently use the vLLM-compatible base and a training framework like Hugging Face's SFTTrainer.
What happens if a document is very long or has many pages?
Process one page (or one image) per inference call. vLLM batching lets you queue multiple pages and parallelize. No multi-page context window; treat each page independently.
Build a Private Document Intelligence System
LightOnOCR-1B is your perception layer. Pair it with LLM.co's operational AI platform to automate invoice processing, compliance triage, and knowledge ingestion—without vendor lock-in or data leaving your infrastructure.