LightOnOCR-1B-1025
LightOnOCR-1B is a compact vision-language model (1.16B parameters) specialized in extracting text from documents, PDFs, forms, and tables. It runs 2–5× faster than comparable OCR systems at a fraction of the cost (~$0.01 per 1,000 pages on H100), supports 10+ European languages, and is available under Apache 2.0 license with no usage gating.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | lightonai |
| Parameters | 1.2B |
| Context window | Unknown |
| License | apache-2.0 — OSI-approved |
| Modality / task | image-to-text |
| Gated on HuggingFace | No |
| Downloads | 194.3k |
| Likes | 254 |
| Last updated | 2026-06-29 |
| Source | lightonai/LightOnOCR-1B-1025 |
What LightOnOCR-1B-1025 is
End-to-end differentiable vision-language model combining a Pixtral-based vision encoder with a Qwen3-based text decoder. Trained on diverse PDF corpora (scientific papers, receipts, invoices, forms, handwritten text). Supports inference via vLLM (v0.11.1+) and native Transformers. Full BF16 precision model (this variant, -1025). Input: single image per inference (1,540px target longest dimension); output: structured OCR text with layout awareness. Multiple vocabulary variants available (32k, 16k tokens) for European language optimization.
Run LightOnOCR-1B-1025 locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="lightonai/LightOnOCR-1B-1025")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
Minimum ~2.3 GB VRAM for BF16 inference (1.16B parameters × 2 bytes/param). Recommended: GPU with 8–16 GB VRAM (e.g., RTX 4070, A30) for comfortable batching and throughput. CPU inference possible but slow (not recommended for production). Paper reports 5.71 pages/s on H100 (80 GB). Estimated 1–2 pages/s on mid-range GPUs; exact VRAM/latency requires profiling on target hardware.
Model is fully differentiable and supports LoRA and full parameter fine-tuning. Official Colab notebook provided for fine-tuning (link in card). Domain adaptation examples: receipts, invoices, scientific papers, forms. Transformers integration for training listed as 'coming soon'; vLLM inference already stable. No QLoRA benchmarks provided; requires testing for memory-constrained fine-tuning.
When to avoid it — and what to weigh
- Real-time, sub-100ms latency requirements — Model inference typically takes seconds per page on consumer GPUs. Suitable for batch processing, not real-time streaming or interactive applications.
- Handwritten or heavily degraded document images — While trained on some handwritten text, no benchmarks provided for severely degraded, cursive, or non-Latin scripts. Test on target data before deployment.
- Ultra-constrained edge deployment (mobile, IoT, sub-1GB RAM) — 1.16B parameters in BF16 precision requires ~2.3GB VRAM minimum. Pruned 32k/16k variants available but performance trade-offs not quantified.
- Proprietary or non-Latin script languages (CJK, Arabic, Indic scripts) — Model optimized for Latin-alphabet European languages. No evaluation data provided for non-Latin scripts; requires custom fine-tuning or alternative.
License & commercial use
Apache License 2.0. Permissive OSI-approved license allowing commercial use, modification, and redistribution with no restrictions on commercial output.
Apache 2.0 is a permissive open-source license with no commercial use restrictions. Model weights and code may be used in commercial products without licensing fees. Training data sourced from diverse PDFs (scientific, synthetic, real documents); dataset to be released under open license per card. No proprietary or restricted training dependencies stated. Commercial deployment requires only standard attribution per Apache 2.0 terms. No gating; no API keys required.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Active |
| Documentation | Strong |
| License clarity | Clear |
| Deployment complexity | Low |
| DEV.co fit | Strong |
| Assessment confidence | High |
No explicit security audit or adversarial robustness claims provided. Model trained on real and synthetic PDFs; potential for latent biases or prompt injection via embedded text in images. Recommend validating OCR output for sensitive compliance use cases (PII extraction, financial documents) and sandboxing in untrusted environments. vLLM deployment should enforce resource limits (--limit-mm-per-prompt) to prevent DoS. No mention of differential privacy or data privacy guarantees in training.
Alternatives to consider
DeepSeekOCR
Comparable OCR vision-language model; LightOnOCR claims 1.73× speedup. Evaluate if DeepSeekOCR offers better multilingual support or domain-specific variants.
PaddleOCR-VL (0.9B)
Lightweight OCR baseline; LightOnOCR claims 2× speedup. PaddleOCR may be preferable if extremely low latency or legacy framework compatibility required.
Qwen-VL or Pixtral (general VLMs)
General-purpose vision-language models may handle broader document types but are larger and slower. Consider if OCR-specific optimization not critical for your use case.
Ship LightOnOCR-1B-1025 with senior software developers
LightOnOCR-1B is production-ready, open-source, and commercially permissive. Use vLLM for high-throughput batch processing or Transformers for fine-tuning on domain-specific documents. No licensing fees, no API limits. Get started with the official Colab notebook or vLLM deployment guide.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
LightOnOCR-1B-1025 FAQ
Can I use LightOnOCR-1B in a commercial product?
What GPU do I need to run this model?
Does it support languages other than European?
Can I fine-tune this model?
Work with a software development agency
Adopting LightOnOCR-1B-1025 is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.
Deploy Fast, Accurate Document OCR Today
LightOnOCR-1B is production-ready, open-source, and commercially permissive. Use vLLM for high-throughput batch processing or Transformers for fine-tuning on domain-specific documents. No licensing fees, no API limits. Get started with the official Colab notebook or vLLM deployment guide.