LightOn

LightOnOCR-3 4B

About

LightOnOCR-3 4B is a 4.5B open-weight model from LightOn. Page to Markdown, or labeled blocks with boxes in one call. It accepts text, images, and documents. Context window is 16K. Released October 6, 2026.

Parameters
4.5B
Published parameter count.
Context window
16K
Tokens of context on a request.
License
Apache-2.0
License on the weights.

Pricing

Token rates are US dollars per 1M tokens. Image rates are per 1K images. Audio rates are per 1M audio seconds.

Rate	Price
Input / 1M	$0.20
Cached input / 1M	$0.20
Output / 1M	$0.50

Compare

LightOnOCR-3 4B is one of the catalog's OCR models. The table is each model's published price, size, and context.

Model	What it does	Price	Parameters	Context
LightOnOCR-3 4B	Page to Markdown, or labeled blocks with boxes in one call.	$0.20 / 1M input	4.5B	16K
PP-OCRv6	PaddleOCR PP-OCRv6 medium text detection and recognition; scene OCR JSONL on image chat; text returns plain text; document_url fan-out via text.	$0.02 / 1M input	20M	—
DeepSeek-OCR-2	DeepSeek OCR 2 and markdown extraction.	$0.25 / 1M input	3.4B	32K
dots.mocr	Multilingual document layout parsing and markdown OCR.	$0.20 / 1M input	3B	32K

Benchmarks

Published scores for LightOnOCR-3 4B, from the LightOnOCR-3 post.

Evals

Reading	Example
Qualitative	Clear structure, grounded in the input
Quantitative	olmOCR-Bench overall: 86.3.
Cost and performance	Lower listed rate, mid-pack latency

Performance

Latency is the end-to-end round trip. Throughput is completion tokens divided by that time. It is a request rate, not decode speed.

Latency
7,925 ms
Throughput
—
Requests
358

Supported inputs

textIgnored
A text part beside the media is accepted and ignored. A request with no media part is a 400.
image_urlYes
One per request. A second image_url part is a 400.
video_urlNo
Not accepted.
document_urlYes
A PDF, up to 128 pages, rasterized per page at document_dpi (default 96). See Document Inputs.

Default method markdown. grounding and layout return the page as blocks, each with a label and a box.
Context window 16,384 tokens, with a maximum output of 4,096 tokens.
Weights: lightonai/LightOnOCR-3-4B, a 4.5B fine-tune of Qwen3.5-4B. Aliases are listed below.

Methods

Methods this model serves. Payload shapes are in the docs.

Method	Returns
markdown	Reading-order Markdown string (default). Tables come back as HTML.
grounding	Block items with `bbox_xywh`, `label`, and `text`
layout	Region items with `bbox_xywh`, `label`, and `order`, no text

Text mode emits the payload alone for one image, and one <document> / <page> block per PDF, each page declaring its kind in format. In JSON mode an image content is the string or the container above, and a document page content is always doc.page.blocks. See Methods & Response Format.

layout is the grounding call with the text dropped, so it bills the same output tokens. grounding writes about 25% more output tokens than markdown. The model grounds on a fixed 0-1000 grid; the Gateway normalizes every box to 0-1 against the image or rasterized page.

Grounding

grounding returns what is on the page and where it is, in one call. Send the media alone; the method picks the prompt:

response = client.chat.completions.create(
    model="lightonai/lightonocr-3-4b",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://storage.googleapis.com/vlm-data-public-prod/hub/examples/document.invoice/invoice_1.jpg"
                    },
                },
            ],
        }
    ],
    extra_body={"method": "grounding"},
)

print(response.choices[0].message.content)

The reply is the block container, alone in text mode and under content in JSON mode. Trimmed to four blocks:

{"object": "doc.page.blocks", "items": [{"block_id": 2, "bbox_xywh": [0.623, 0.112, 0.194, 0.034], "label": "title", "text": "## INVOICE"}, {"block_id": 3, "bbox_xywh": [0.514, 0.187, 0.22, 0.028], "label": "text", "text": "**Date:** November 11, 2023 **Invoice #** 9999999"}, {"block_id": 9, "bbox_xywh": [0.571, 0.639, 0.26, 0.099], "label": "table", "text": "<table><tr><td>Subtotal</td><td>400.00</td></tr></table>", "text_format": "html"}, {"block_id": 10, "bbox_xywh": [0.323, 0.777, 0.275, 0.013], "label": "footer", "text": "THANK YOU FOR YOUR BUSINESS!"}]}

Label	Text
`text`, `title`, `list`, `header`, `footer`, `page_number`, `footnote`, `caption`, `code`, `aside_text`	Markdown
`table`, `chart`	HTML, with `text_format: "html"`. A chart becomes a table of its data points.
`formula`	LaTeX
`image`, `header_image`, `footer_image`	A one-line description of the image

A trailing + on a label marks a block that continues the one before it. On a PDF each page's content is doc.page.blocks with the same items. Set method="layout" for the boxes and labels alone.

Aliases

lightonai/lightonocr-3-4b (preferred), lightonocr-3-4b, and lightonai/LightOnOCR-3-4B all resolve to this model.

Estimate cost

Estimate. A page is 2,500 input tokens and 1,000 output tokens.

Pages

$1.00

Quick start

Call this model on the OpenAI-compatible gateway. The model id is already filled in.

from openai import OpenAI

client = OpenAI(
    base_url="https://gateway.vlm.run/v1/openai",
    api_key="<VLMRUN_API_KEY>",
)

response = client.chat.completions.create(
    model="lightonai/lightonocr-3-4b",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "document_url",
                    "document_url": {
                        "url": "https://storage.googleapis.com/vlm-data-public-prod/hub/examples/finance.sec-filings/tsla-8k.pdf"
                    },
                },
            ],
        }
    ],
    extra_body={"method": "markdown"},
)

print(response.choices[0].message.content)

Response

One <document> block per PDF, one <page> block per page. A failed page is self-closing, with status="error" and no body:

tsla-8k.pdf2 pages

Page 0

Annual Report 2024

Overview

The company reported revenue of $1.2M.

Metric	2023	2024
Revenue	0.9M	1.2M

Page 1 · failed

This page returned no body.

A single image returns the Markdown string alone, with no wrapper.