LightOnOCR-3 (Document OCR)
LightOnOCR-3 is LightOn's open-weight OCR model (Apache-2.0, released October 2026). It reads a page image and returns markdown. With grounding turned on, it also returns a label and a bounding box for every block, and renders tables and charts as HTML tables. OpenTryOn exposes it on the parse service, next to Nemotron Parse.
Fashion uses include size charts, spec and tech sheets, line sheets, invoices and care-label scans.
| Registry id | How it runs | Needs |
|---|---|---|
lightonocr-3-local | In-process Hugging Face Transformers | GPU or Apple Silicon; transformers>=5.2, torch |
lightonocr-3-vllm | Client for a vLLM / SGLang server you run | Your server; nothing heavy on the client |
LightOn has no hosted API for LightOnOCR-3, and no third-party host serves it.
Variants
--variant | Model | Notes |
|---|---|---|
0.8b (default) | lightonai/LightOnOCR-3-0.8B | About 0.9B parameters, roughly 2 GB in BF16 |
4b | lightonai/LightOnOCR-3-4B | LightOn's recommended quality tier |
Both use the Qwen3.5 vision-language architecture. The 1B sibling keeps the older LightOnOCR-2 architecture and isn't wired here. LightOn reports 85.5 (0.8B) and 86.3 (4B) on olmOCR-Bench; these are vendor figures.
CLI
# Markdown from one page image
opentryon parse --model lightonocr-3-local --image size_chart.png
# Labelled blocks with bounding boxes, pages 1-2 of a PDF, larger model
opentryon parse --model lightonocr-3-local --image tech_pack.pdf --pages 1-2 --grounding --variant 4b
# Against your own vLLM server
vllm serve lightonai/LightOnOCR-3-0.8B --trust-remote-code \
--default-chat-template-kwargs '{"enable_thinking": false}'
opentryon parse --model lightonocr-3-vllm --image size_chart.pdf --grounding
| Flag | Meaning |
|---|---|
--image, -i | Page image or PDF (path or URL) |
--grounding | Also return labelled blocks with bounding boxes |
--pages | PDF pages to read, e.g. 1-3,5 (default: all) |
--max-tokens | Max tokens per page (default 4096) |
--variant, --model-id, --device | lightonocr-3-local only |
--base-url, --served-model | lightonocr-3-vllm only (defaults from LIGHTONOCR_BASE_URL, LIGHTONOCR_MODEL) |
PDFs are rendered with pypdfium2 (pip install pypdfium2) so the longest side is 2048 px, LightOn's recommended setting.
Output
The result is saved as JSON:
{
"model": "lightonai/LightOnOCR-3-0.8B",
"grounding": true,
"page_count": 2,
"markdown": "## Size guide\n\n<table>...</table>",
"blocks": [
{"page": 1, "label": "title", "bbox": [113, 608, 263, 623], "text": "## Size guide", "continues": false}
],
"pages": [{"page": 1, "markdown": "...", "blocks": ["..."], "raw": "..."}]
}
bboxis[x1, y1, x2, y2], normalised to 0–1000 on both axes, whatever the image size.- Labels include
text,title,list,table,chart,image,caption,formula,code,header,footerandpage_number. continues: truemarks a block that carries on from the previous one, for example across columns;markdownjoins them.- Without
--grounding, each page is plain markdown andblocksis empty.
Python
from tryon.models.lightonocr import LightOnOCRLocalAdapter
from tryon.api.lighton import LightOnOCRServerAdapter
result = LightOnOCRLocalAdapter(variant="0.8b").parse_document("size_chart.pdf", grounding=True)
print(result["markdown"])
served = LightOnOCRServerAdapter(base_url="http://127.0.0.1:8000/v1")
print(served.parse_document("care_label.jpg")["markdown"])
Requirements
lightonocr-3-local loads with Qwen3_5ForConditionalGeneration, which needs transformers>=5.2 (the model card was saved with 5.5.4). That's newer than the shared opentryon[local] pin, so use a separate environment if loading fails:
pip install "transformers>=5.5.4" torch pillow pypdfium2
See also: Nemotron Parse (hosted NIM OCR), DeepSeek-OCR.