LightOn releases LightOnOCR-3, open OCR models that also turn charts into tables

French AI company LightOn released the LightOnOCR-3 model family on October 8, 2026. OCR (optical character recognition) models turn the text in a scanned or photographed document into editable text. LightOnOCR-3 comes in three sizes, 0.8B, 1B and 4B, and is released under Apache 2.0 for both research and commercial use. The weights are openly available, so organizations can run the models on their own servers.

According to the company's announcement, this generation brings three new abilities. The models can mark every block on a page with a labeled box that says what it is. They write a short description, about ten words at most, for images that are not charts. And they extract the data in each chart as an HTML table. With an empty prompt the model only transcribes the text; with the grounding prompt it also adds the page layout, image descriptions and chart tables.

Qwen3.5 architecture and benchmark results

The 0.8B and 4B models use the Qwen3.5 vision-language architecture; according to the announcement, the 4B model is based on Qwen3.5-4B. The 1B model keeps the previous generation's architecture. In LightOn's results, the 4B model scores 86.3 on olmOCR-Bench; the 35.1-billion-parameter Infinity Parser Pro tops the table at 87.6, while Chandra 2, which is the same size, scores 85.8. On ParseBench, which covers five categories, the 4B and 0.8B models take the top two places with 75.1 and 74.6. The 4B model also leads the chart reading and semantic formatting categories.

Speed was measured on a single H100 GPU with vLLM. According to the company, at 1540 pixels the 0.8B model reaches 4.78 pages per second and the 4B model 3.36; the 1B model has the lowest single-page latency at 2.7 seconds. LightOn published scripts on GitHub for converting the raw output, visualizing it and reproducing the benchmarks, and there is a demo on Hugging Face.

The announcement says the training data included synthetic pages in more than 20 languages, but it gives no separate results for languages such as Turkish or Arabic; a separate result was shared for French documents. It is worth testing the models on your own documents first. LightOn ran the comparisons, and rival models lead in some categories. For a fuller picture, follow evaluations from independent researchers as well.

Sources: LightOn: LightOnOCR-3 announcement, Hugging Face: LightOnOCR-3-1B model card, GitHub: LightOnOCR.