What is LightOnOCR-3-4B?
LightOnOCR-3-4B is the largest of three OCR models LightOn, a Paris AI company, released on 8 October 2026 under Apache 2.0. It is built on the Qwen3.5 vision-language architecture; LightOn calls it 4B, and the checkpoint Hugging Face counts holds 4.54 billion parameters. The family also has a 0.8B on the same architecture and a 1B that keeps the older LightOnOCR-2 design as a drop-in upgrade.
Called with an empty prompt, it transcribes the page to Markdown as earlier versions did. Called with the prompt "grounding", it labels each block with one of 16 types and a bounding box, writes a short description of each image, and turns charts into an HTML table of their data points, for about 25% more output tokens.
It replaces LightOnOCR-2-1B, released in January 2026, which drew 779,954 downloads in the past month on Hugging Face against 19,097 for the 4B. That model was transcription only, ran at 5.71 pages a second on one H100 by LightOn's count, and scored 0.0 on ParseBench's visual-grounding category.
Is LightOnOCR-3-4B worth it?
LightOnOCR-3-4B is the most complete end-to-end OCR model in its size class, on LightOn's own benchmarks. It scores 75.1 overall on ParseBench, ahead of Infinity-Parser2-Pro at 74.3 with a ninth of the parameters, Chandra 2 at 70.1 and Mistral OCR 4.1 at 68.2, and 86.3 on olmOCR-Bench, 1.3 points behind Infinity-Parser2-Pro and 0.5 ahead of Chandra 2. Every figure is LightOn's run, after normalisation functions it publishes along with scripts to reproduce the results.
The French-document results are the most distinctive. On fr-bench-pdf2md it scores 74.1 overall against 69.0 for Chandra 2 and 63.2 for Infinity-Parser2-Pro, with 46.1 on handwritten pages where Chandra 2 has 26.1 and Mistral OCR 4.1 has 11.5. For European archives and administrations that is a larger gap than any single English benchmark shows.
There is no OmniDocBench score, so it cannot be placed directly against PaddleOCR-VL-1.6, OvisOCR2 or GLM-OCR on the most common table. Its weak category on olmOCR-Bench is old scans at 52.5, where Infinity-Parser2-Pro reaches 58.2, and the 0.8B sibling comes within half a point of it on ParseBench at a fifth of the size. Only two prompts are supported; LightOn says any other instruction is out of distribution.
Speed is measured carefully. On one H100 with vLLM, it serves 3.36 pages a second at peak when pages are rendered at 1,540 pixels on the long side, 66% more than at the 400 DPI setting behind its best scores. Against Chandra-OCR-2, built on the same Qwen3.5-4B base, it is 19% faster on a single page and serves 21% more pages a second, mainly because it writes 14% fewer tokens.
Our take: LightOnOCR-3-4B is the European counterpart to the Chinese small parsers, with a broader output than most of them. The 0.8B variant is the better default where throughput matters more than handwriting, since it gives up only half a ParseBench point.
Where it earns its place
A records team at a French local authority digitising handwritten and typed case files
Turn mixed handwritten and printed pages into searchable text on in-house servers
Serve the 4B with vLLM on one H100 and send pages rendered at 400 DPI. LightOn's 46.1 on handwritten French pages is well ahead of rivals it tested, and the Apache 2.0 licence suits public procurement.
An ML engineer building a RAG system over corporate reports
Chunk documents by real sections and keep chart values searchable
Call with the grounding prompt so each block comes back labelled with a box, and index chart tables as data. Grounding adds about 25% output tokens, roughly 1,433 instead of 1,158 per page.
A platform team already running LightOnOCR-2-1B in production
Upgrade OCR without changing their client code
Swap in LightOnOCR-3-1B, which keeps the old architecture, or move to the 4B on Transformers 5.2 or later; the empty-prompt transcription mode behaves as before.
Before you commit
- Setup
- pip install transformers 5.5.4 or later with pillow and pypdfium2 and load it with the Qwen3.5 classes, or vllm serve it with thinking disabled; render PDFs at 400 DPI with a 2,048-pixel longest edge.
- Integrations
- Transformers and vLLM; LightOn's GitHub repository adds a client, CLI, output converters, a viewer and benchmark reproduction code; community GGUF builds on Hugging Face.
- Data and privacy
- Apache 2.0 for research and commercial use. Self-hosted, documents stay on your hardware; the Hugging Face demo is the only hosted option named on the card.
- Cost at scale
- The BF16 checkpoint is about 9 GB, so one data-centre GPU serves it. LightOn reports a peak of 3.36 pages a second on one H100 at 1,540-pixel rendering, roughly 12,000 pages an hour.
- Limits
- Only the empty and grounding prompts; trained and evaluated with thinking off; old scans are its weakest olmOCR-Bench category; no OmniDocBench results published.
What’s good
- 75.1 on ParseBench and 86.3 on olmOCR-Bench (self-reported)
- Layout boxes, image descriptions and chart tables in one pass
- 74.1 on French documents, 46.1 on handwritten pages
- Apache 2.0 from a European maker, with reproduction scripts
What’s not
- No OmniDocBench figure to compare with Chinese rivals
- All benchmark numbers are LightOn's own runs
- Only an empty prompt or "grounding" is supported
- The 0.8B sibling is nearly as accurate on ParseBench
Best for
Teams parsing European-language business documents, forms and handwriting who want layout and chart data without running a detection pipeline.
LightOnOCR-3-4B alternatives
LightOnOCR-3-4B: common questions
- Is LightOnOCR-3-4B worth it?
- LightOnOCR-3-4B gives one 4B model the layout boxes and chart data that used to need a pipeline, and on LightOn's numbers it is the open model to beat on French and handwritten documents. We rate it 4.2 out of 5.
- How much does LightOnOCR-3-4B cost?
- Free under Apache 2.0. No first-party hosted API listed on the model card; a free demo runs on Hugging Face Spaces.
- What are the downsides of LightOnOCR-3-4B?
- No OmniDocBench figure to compare with Chinese rivals; all benchmark numbers are LightOn's own runs; only an empty prompt or "grounding" is supported; the 0.8B sibling is nearly as accurate on ParseBench.
- Who is LightOnOCR-3-4B best for?
- Teams parsing European-language business documents, forms and handwriting who want layout and chart data without running a detection pipeline.
- How much does LightOnOCR-3-4B cost for a team or in production?
- The BF16 checkpoint is about 9 GB, so one data-centre GPU serves it. LightOn reports a peak of 3.36 pages a second on one H100 at 1,540-pixel rendering, roughly 12,000 pages an hour.
- Is LightOnOCR-3-4B safe for company data?
- Apache 2.0 for research and commercial use. Self-hosted, documents stay on your hardware; the Hugging Face demo is the only hosted option named on the card.
- What are the best LightOnOCR-3-4B alternatives?
- The closest alternatives we have reviewed are Chandra OCR 2, Infinity-Parser2-Flash, Mistral OCR 4.1 and Surya OCR 2.5.
Answers checked Nov 2, 2026.
Make LightOnOCR-3-4B? Show this rating on your site
