Convert pages to Markdown 48% cheaper than Mistral OCR 4.1

LightOnOCR-3 turns scanned page images into Markdown with table structure and reading order. SIE is the open-source inference engine that serves this document model.

Get started Run in your cloud

LightOnOCR-3 on SIE

Table rows kept together

Page image

Crop of the Inti Everspring, Methomyl 90 SP technical data sheet: a borderless specification table with characteristic, specification and method columns. — Inti Everspring, Methomyl 90 SP technical data sheet

Markdown

Characteristic	Specification	Method of analysis
Al, as Methomyl	87.5 - 92.5 %	High Performance Liquid Chromatography
pH (1 % in water)	4.0 - 8.0	pH-meter
Appearance	White to cream powder	Visual

Missing tables and altered words break the record

GPT-6 Luna (no reasoning)

Title word changed

Source title

TÍTULO DEL DOCUMENTO

El ciclomontañismo una experiencia significativa Universitaria, en la Universidad de Cundinamarca

Universidad de Cundinamarca, document title

Returned Markdown (excerpt)

El ciclomañanismo una experiencia significativa Universitaria, en la Universidad de Cundinamarca

The title’s key word changes. An exact search for the original title no longer matches.

Read the complete recording

Claude Sonnet 5.5 (thinking off)

Metric tables omitted

One of the source’s metric tables

Source table: accuracy 95%, sensitivity 93.75%, specificity 96.25%, precision 96.15%, F1 score 0.9494 and MCC 0.9003. — Xu et al., Scientific Reports (2022), Figure 1A

Returned Markdown (excerpt)

**Figure 1.** The classifier of the DNN models based on the matrices $M$, $M_{kinematics}$, $M_{kinetics}$.

All three metric tables are absent from the complete response.

The caption survives; the values needed to compare the results disappear.

Read the complete recording

arXiv:2604.08538 Study of 14 document parsers on 2,000 business pages: none is consistently strong across all five areas it tests
ParseBench: A Document Parsing Benchmark for AI Agents

Boyang Zhang, Sebastián G. Acosta, Preston Carlson and others

“no method is consistently strong across all five dimensions”
9 April 2026
Microsoft Foundry Blog Microsoft: when a model loses table structure, teams bolt on a layout parser
“When the model does not preserve table structure, you add a layout parser.”
29 September 2026

Convert pages 90% cheaper than GPT-6.1 Sol

Run this eval

Document checks passed

8,234 checks ↑

GPT-6.1 Sol (medium reasoning)

GPT-6.1 Sol (medium reasoning) 84.6%

Claude Sonnet 5.5

Claude Sonnet 5.5 84.2%

Mistral OCR 4.1

Mistral OCR 4.1 83.3%

GPT-6 Luna

GPT-6 Luna 80.7%

Claude Haiku 5.5

Claude Haiku 5.5 80.1%

SIE LightOnOCR-3

SIE LightOnOCR-3 73.2%

Cost per 1,000 pages

USD ↓

GPT-6 Luna

$1.01 GPT-6 Luna

Claude Haiku 5.5

$1.24 Claude Haiku 5.5

SIE LightOnOCR-3

$2.05 SIE LightOnOCR-3

Mistral OCR 4.1

$4.00 Mistral OCR 4.1

GPT-6.1 Sol (medium reasoning)

GPT-6.1 Sol (medium reasoning) $21.79

Claude Sonnet 5.5

Claude Sonnet 5.5 $25.42

Evaluation

Tested: olmOCR-Bench: 1,366 pages across 1,068 source families
Scored: Equal-weight text, reading-order, table and math checks across eight categories
Prices: SIE Serverless; provider list prices on recorded usage

Send a page image, get Markdown back

The response is Markdown with HTML tables and image placeholders in entities[0].text. The preview renders the recorded table.

from urllib.request import urlopen
from sie_sdk import SIEClient
client = SIEClient(api_key="API keysk-sie-…")
image_bytes = urlopen(imagemethomyl-page.pngbrowse).read()
image = {"data": image_bytes, "format": "png"}
result = client.extract(
    "modellightonai/LightOnOCR-3-4B",
    {"id": "methomyl-page", "images": [image]},
    options={"max_new_tokens": 12288, "temperature": 0.1, "top_p": 1},
)
print(result["entities"][0]["text"])

import { unpackMessage } from '@superlinked/sie-sdk';

const imageResponse = await fetch("https://superlinked.com/reference/doc-to-markdown/methomyl-page.png");
if (!imageResponse.ok) throw new Error(`Image returned ${imageResponse.status}`);
const image = new Uint8Array(await imageResponse.arrayBuffer());
const response = await fetch("https://api.superlinked.com/v1/extract/lightonai%2FLightOnOCR-3-4B", {
  method: 'POST',
  headers: {
    Authorization: "Bearer sk-sie-…",
    'Content-Type': 'application/json',
    Accept: 'application/msgpack',
  },
  body: JSON.stringify({
    items: [{ id: "methomyl-page", images: [{ data: Buffer.from(image).toString('base64'), format: "png" }] }],
    params: {"options":{"max_new_tokens":12288,"temperature":0.1,"top_p":1}},
  }),
});
if (!response.ok) throw new Error(`SIE returned ${response.status}`);
const result = unpackMessage<{ items: { id: string; entities: { text: string }[] }[] }>(
  new Uint8Array(await response.arrayBuffer()),
);
const item = result.items[0];
if (result.items.length !== 1 || item?.id !== "methomyl-page" || !item.entities[0]?.text) {
  throw new Error('Invalid OCR response');
}
console.log(item.entities[0].text);

curl --fail --silent --show-error 'https://superlinked.com/reference/doc-to-markdown/methomyl-page.png' -o 'methomyl-page.png'
image_bytes=$(base64 < 'methomyl-page.png' | tr -d '\n')
curl https://api.superlinked.com/v1/extract/lightonai%2FLightOnOCR-3-4B \
  -H "Authorization: Bearer sk-sie-…" \
  -H "Content-Type: application/json" \
  -d "{\"items\":[{\"id\":\"methomyl-page\",\"images\":[{\"data\":\"$image_bytes\",\"format\":\"png\"}]}],\"params\":{\"options\":{\"max_new_tokens\":12288,\"temperature\":0.1,\"top_p\":1}}}"

Build the "Doc to Markdown" capability into my app using the Superlinked Inference Engine (SIE).

Context
- SIE is an OpenAI-style inference API. Python SDK: `from sie_sdk import SIEClient`; TypeScript: `@superlinked/sie-sdk`.
- Base URL: https://api.superlinked.com (or my regional endpoint). Auth: Bearer key from env `SIE_API_KEY` (never hard-code it).
- Model: lightonai/LightOnOCR-3-4B (SIE primitive: /extract). Keep the model id configurable.

Task
- Input: an uploaded image containing text.
- Behaviour: return the text visible in the image
- Fetch the unmodified PNG from https://superlinked.com/reference/doc-to-markdown/methomyl-page.png. Send one POST https://api.superlinked.com/v1/extract/lightonai%2FLightOnOCR-3-4B with item id "methomyl-page" and images [{data: original PNG bytes, format: "png"}]. Preserve params {"options":{"max_new_tokens":12288,"temperature":0.1,"top_p":1}}; keep this model and profile selected and fail when unavailable instead of substituting another model. In Python use SIEClient.extract; TypeScript uses the native REST body and public unpackMessage decoder shown in Copy. Read the transcription from items[0].entities[0].text (Python single-item result["entities"][0]["text"]), without repairing or calculating source values.

Deliverables
- A typed client wrapper, an application-level function for this task, error handling for timeouts/empty input, and unit tests with a stubbed client.
- Wire it into my existing stack (ask me which framework if unclear) and add a short usage example.

Run

Table in model response

METHOMYL 90 SP

Characteristic	Specification	Method of analysis
Al, as Methomyl	87.5 - 92.5 %	High Performance Liquid Chromatography
pH (1 % in water)	4.0 - 8.0	pH-meter
Appearance	White to cream powder	Visual

Convert a million pages for $2,050

Serverless API

We host the models; you call one API.

No idle GPUs, pay for what you use
Fits your stack: SDK, API, CLI, MCP
Self-host the same open-source engine
SOC 2 Type 2, US or EU data residency
Create an account to try Serverless

Get started

Self-hosted (Kubernetes)

Your data stays in your cloud, backed by our engineers.

Terraform modules for your cloud
Apache-2.0, same engine as Serverless
GPU workers scale to zero between jobs
Separate worker pools for each tenant

Agent prompt Deploy SIE to our AWS account with the superlinked/sie/aws Terraform module. Docs: superlinked.com/docs/deployment.

Deploy SIE to our Google Cloud project with the superlinked/sie/google Terraform module. Docs: superlinked.com/docs/deployment.

Deploy SIE to our Azure subscription with the superlinked/sie/azure Terraform module. Docs: superlinked.com/docs/deployment.

Deploy SIE to our Alibaba Cloud account with the superlinked/sie/alicloud Terraform module. Docs: superlinked.com/docs/deployment.

Talk to us Deploy guide

Run locally

The open-source server on your own machine.

NVIDIA GPU
One container, no cluster
Process your data fully offline
Same SDK and model IDs as Serverless

docker run --gpus all -p 8080:8080 ghcr.io/superlinked/sie-server:latest-cuda12-sglang-vision-extract

Quickstart

Switch to SIE in 5 minutes

from mistralai.client import Mistral

page_url = "https://example.com/scanned-page.png"
client = Mistral(api_key="…")
result = client.ocr.process(
    model="mistral-ocr-4-1",
    document={"type": "image_url", "image_url": page_url},
)
print(result.pages[0].markdown)

SIE

from sie_sdk import SIEClient

page_url = "https://example.com/scanned-page.png"
client = SIEClient(api_key="sk-sie-…")
result = client.extract(
    "lightonai/LightOnOCR-3-4B",
    {"images": [page_url]},
    options={"max_new_tokens": 12288, "temperature": 0.1, "top_p": 1},
)
print(result["entities"][0]["text"])