Open-Weight LLM · Private & Custom AI
LightOnOCR-2-1B
A 1B vision-language model for document OCR and text extraction that runs efficiently on modest hardware—purpose-built for companies automating document processing at scale within their own infrastructure.
LightOnOCR-2-1B is a lightweight end-to-end vision-language model that converts PDFs, scans, and images into structured text without external OCR pipelines. For ops teams, it handles tables, forms, receipts, and multi-language documents; deployed privately, it keeps sensitive document data in your environment while processing ~500k pages/day per H100 at <$0.01 per 1k pages.
Model facts
Private deployment
Run LightOnOCR-2-1B in your own environment
The model fits in ~4–6 GB VRAM (bfloat16 estimate) and runs on single GPUs or CPU inference. Self-hosting means document data never leaves your infrastructure—critical for financial records, contracts, medical files, or compliance-sensitive materials. Transformers 5.0+ and vLLM both support it; deployment to Kubernetes, Azure, or on-prem is straightforward. You own the inference pipeline, control versioning, and avoid vendor lock-in on document processing.
Operational AI use cases
Automated Invoice & Receipt Processing
Route incoming invoices (email, scans, PDFs) through LightOnOCR privately to extract line items, amounts, vendor details, and tax codes. Feed structured output to accounting systems (NetSuite, SAP, custom ledgers) via API. Reduces manual data entry; keeps vendor/payment data internal.
Internal Knowledge Base Ingestion
Scan internal policy documents, training manuals, regulatory guides, archived reports into your private LLM knowledge layer. OCR captures tables, diagrams, multi-column layouts—then embed and index for agent-driven Q&A. Data stays within your security boundary; no cloud document storage.
Contract & Legal Document Triage
Extract clauses, dates, party names, payment terms from incoming contracts or NDAs without exposing them to third-party services. Classify by risk level (auto-flag unusual terms) and route to appropriate teams. Fully private; audit-trail stays internal.
Custom AI
As a base for custom AI
Strong foundation for domain-specific document AI applications. The base variant supports LoRA fine-tuning on specialized corpora (e.g., medical records, insurance claims, manufacturing specs). Fully differentiable pipeline means you can adapt it to custom output formats (structured JSON, domain ontologies) and integrate it into a private AI product—a document understanding layer for custom chatbots, workflow automation, or vertical SaaS.
In the operating system
Where it fits
Sits in the **perception/ingestion layer** of an AI operating system: it bridges unstructured documents (images, PDFs, scans) into clean, semantically rich text that feeds downstream agents, workflows, and knowledge retrieval. Use it as the first node in document-centric workflows—transcription → classification → extraction → knowledge indexing.
Data control & security
Self-hosting eliminates third-party document transmission; sensitive PDFs, contracts, invoices, and forms stay in your data center or private cloud. No cloud logging, no model training on your data. Audit and compliance teams see the entire inference stack. Note: security posture depends on your infrastructure (network isolation, encryption at rest/in-flight, access controls)—the model itself is neutral; deployment architecture determines data control.
Hardware footprint
**Estimate (bfloat16):** ~4–6 GB VRAM on H100/A100. **fp32:** ~8–10 GB. Supports CPU inference with quantization (int8, int4) for latency-tolerant workloads. Processes ~5.71 pages/sec on single H100; batch processing on multi-GPU setups scales linearly.
Integration
Transformers 5.0+ integration is native (LightOnOcrProcessor, LightOnOcrForConditionalGeneration). vLLM support for batched inference. Standard JSON/REST APIs for document submission. Pair with document queuing (RabbitMQ, Kafka), database staging (PostgreSQL for extracted metadata), and webhook triggers to downstream systems (accounting, CRM, knowledge store). Render PDFs at 200 DPI, target longest edge ~1540px for optimal performance.
When it's not the right fit
- —Extreme real-time latency required (<100ms end-to-end): model inference adds 150–300ms per page.
- —Handwritten or heavily degraded documents: model trained on printed/digital text; cursive/faded scans degrade accuracy.
- —Extremely niche languages or specialized notation not in training corpus (e.g., rare indigenous scripts, proprietary symbols): multilingual coverage is strong (11+ languages) but not exhaustive.
- —Sub-page-level structured extraction without fine-tuning: base model outputs clean text; complex hierarchical extraction (exact cell coordinates, nested tables) requires domain adaptation.
Alternatives to consider
Chandra-1B (Deci AI)
Comparable 1B parameter model; claims faster inference on specific hardware. Less multilingual coverage; OCR benchmarks trail LightOnOCR-2 on OlmOCR-Bench.
OlmOCR (AI2)
Larger, open-weight model with strong academic OCR results. ~3.3× slower than LightOnOCR-2; higher VRAM footprint. Better for high-accuracy offline batch pipelines.
Tesseract 5 + LayoutLM (Open source classical + transformers hybrid)
Modular, no neural network inference for base OCR; lightweight. Harder to handle complex layouts, math, tables; requires manual pipeline tuning; no end-to-end differentiability.
Related open models
FAQ
Can we fine-tune LightOnOCR-2 on our proprietary documents?
Yes. The base variant (LightOnOCR-2-1B-base) is provided for fine-tuning. It's fully differentiable; use LoRA for parameter-efficient adaptation to domain-specific layouts (receipts, forms, legal docs). A Colab notebook is linked in the model card. Fine-tuning stays private—your data and checkpoints remain internal.
What's the commercial/licensing story for internal ops use?
Apache 2.0 license: permissive, allows commercial use, internal deployment, and modifications without restriction. No usage fees, no licensing checks. You can redistribute it if you include a copy of the license.
How do we deploy this privately without exposing documents?
Deploy via Transformers or vLLM on your own GPU cluster, Kubernetes, or on-prem hardware. Documents are submitted via local API calls; inference runs entirely in your environment. Use network isolation (private subnets, VPN) and encryption for data at rest/in-flight. The model has no built-in telemetry or external dependencies.
How does accuracy compare to cloud-based OCR services?
On standard OCR benchmarks (OlmOCR-Bench), LightOnOCR-2-1B outperforms or matches larger competitors while being ~9× smaller and faster. However, benchmarks are public datasets; your specific domain (handwriting, ultra-low resolution, niche layouts) may differ. Test on your corpus before full rollout; fine-tune if needed.
Build a Private Document AI System
LightOnOCR-2 is ready to deploy in your own infrastructure. Pair it with LLM.co's workflow and agent layers to automate invoicing, contracts, and internal knowledge ingestion—all with data staying in-house. Let's design your custom ops AI stack.