LightOn released three LightOnOCR-3 weights under Apache 2.0 with grounding outputs and benchmark reproduction scripts.
LightOn releases OCR models that extract layout, charts and images
2026-10-08Verified announcementTRIAL
LightOn released 0.8B, 1B and 4B LightOnOCR-3 weights. The models convert pages to Markdown; a grounding mode also emits labeled bounding boxes, image descriptions and chart data. The Apache 2.0 weights permit commercial use. Reproduction scripts are public, but the reported scores remain vendor-run.
What this means for you
For RAG and document-pipeline teams, one model handling transcription, layout grounding and chart extraction could replace several OCR and layout stages. The vendor also shows that post-processing and resolution materially affect benchmark scores, so teams must retest on their own scans, tables and handwriting.
Evidence & related material
Technical details & scores
| Models | 0.8B, 1B and 4B variants; 0.8B and 4B use Qwen3.5 vision-language backbones |
|---|---|
| Outputs | Markdown transcription; grounding mode adds labeled boxes, image descriptions and chart tables |
| Vendor Benchmarks | 4B: olmOCR-Bench 86.3, ParseBench 75.1 and fr-bench-pdf2md 74.1 |
| Serving Test | 512 pages on one H100 with vLLM 0.30.0; resolution changes throughput materially |
| License | Apache 2.0 |
| Artifacts | weights, conversion and visualization code, benchmark reproduction scripts and demo are public |
| Time Basis | Hugging Face article provides date-only publication metadata: 2026-10-08; no time of day was published |
Technical impact 4
Engineering value 5
Adoption signal 2
Maturity 3
Verification cost 2
Risk 2