Tasks and inputs
LightOnOCR-3-0.8B is listed for Image-text-to-text, Image-to-text, Document OCR. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
LightOnOCR-3-0.8B is LightOn's smallest open document model: page images in, markdown out, with optional bounding boxes and chart tables.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Free open weights; a registry lists a hosted price of $2.464/1M input that was not checked.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Send LightOnOCR-3-0.8B an empty prompt and a page image, and you get the whole page back as markdown. Send it the word "grounding" instead, and it adds labelled bounding boxes, short descriptions of embedded images, and chart data rendered as HTML tables. Two modes, one tiny model.
It is the smallest of LightOn's three LightOnOCR-3 document models, released on 8 October 2026 alongside 1B and 4B versions under Apache-2.0.
The 0.8B is not just a shrunk 1B. It uses the Qwen3.5 vision-language architecture, while the 1B keeps the LightOnOCR-2 design. If you benchmarked the previous generation, this model may behave differently in ways that are not just about size.
How small is it? The launch materials call it 0.8B, the Hugging Face page shows 0.9B, and the safetensors total about 853M parameters. Any of those numbers means it runs on modest hardware.
LightOn reports up to 4.78 pages per second on a single H100 with vLLM, when pages are rendered at 1540 pixels. Rendering resolution matters a lot here. Go much lower and small print suffers; go higher and throughput drops.
| Item | Detail |
|---|---|
| Hugging Face | lightonai/LightOnOCR-3-0.8B |
| Parameters | ~0.85B |
| Licence | Apache-2.0 |
| olmOCR-Bench | 85.5 |
| Long tiny-text category | 94.1, the category leader |
| Thinking | Trained and evaluated with it off |
| Transformers | 5.2 or newer |
Those benchmark scores are from LightOn's published results.
A registry lists a 262,144-token context and 16,384-token maximum output. Neither figure appears in LightOn's materials, so treat them as unverified. A page of dense text rarely approaches either limit anyway.
The model was trained and evaluated with thinking disabled. The Qwen3.5 base may expose a thinking switch, but enabling it puts the model outside what it was tuned for. Leave it off.
Pull fifty real pages from your archive: scanned invoices, a two-column paper, a table that spans a page break, something with footnotes in 7-point type. Render at 1540 px, run both modes, and diff the markdown against a hand-checked version. Pay attention to table structure and reading order, the two places OCR models most often go wrong.
If 0.8B struggles, try the 4B sibling before switching vendors. For other options see document OCR models, compare with Mistral OCR 4.1, or browse the LightOn provider page.
Compare the model on the work you need to complete.
LightOnOCR-3-0.8B is listed for Image-text-to-text, Image-to-text, Document OCR. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
OpenBMB
MiniCPM-OCR 1.0 is OpenBMB's open 1.3B document-parsing model for text, tables, formulas and charts, with a free API and on-device apps.
OpenBMB
MiniCPM-OCR Jev is OpenBMB's open 2B judge that scores whether an OCR result for a text, formula or table region is correct, from 0 to 1.
Tencent
Youtu-Parsing-Omni is Tencent Youtu's 5B open parser that turns documents, charts, images, audio and video into one structured JSON format.
Sarvam
Sarvam Vision 2.1 is Sarvam's 3B document-intelligence VLM for OCR, tables, forms and Indic handwriting in 22 Indian languages plus English.
NVIDIA
Nemotron 3.5 Parse is NVIDIA's 0.631B open document parser that turns page images into Markdown, HTML or JSON with tables, charts and reading order.
Cohere
Cohere Parse (parse-v5.0) is a 2.3B vision-language document parser that turns PDFs, slides and scans into Markdown for RAG, at $1.50 per 1,000 pages.