Rune-1 OCR models for Arabic, Indic, CJK and historical scripts.

Rune-1 is Runemic's OCR model family. It finds the layout, works out the reading order, then reads each line into Markdown or JSON.

In training Meanwhile, try the preview models free.

  1. 1Layout
  2. 2Reading order
  3. 3Recognition
  4. 4Structure
A typed German letter about digitizing church registers
page.md

      

A sample letter, shown step by step. Rune-1 is in training; the preview models read pages like this today.

Rune-1 comes in three versions: Small, Rune-1 and Custom.

All three are in training. Rune-1 Small will be open and run on your own hardware. Rune-1 is the larger model behind the API. Rune-1 Custom is trained for a particular script or collection.

Open weights, planned

Rune-1 Small

A compact model for local and on-device use. It will be published on Hugging Face under an open licence when it is ready.

  • Multilingual and multi-script
  • Made to run on a single GPU, or a laptop
  • Text lines with layout blocks
  • Fine-tune on your own pages
Follow on Hugging Face
Enterprise

Rune-1 Custom

Rune-1 adapted to a script, historical hand or document type that existing OCR handles badly, trained on your material.

  • Rare and historical scripts
  • Domain forms and ledgers
  • Private weights, your servers
Talk to us

Arabic, Devanagari, CJK and more: scripts that most OCR handles badly.

Most OCR is tuned for Latin text in neat columns. We train Rune-1 on right-to-left, vertical and mixed-script pages from the start: Arabic, Persian (Farsi) and Urdu, including nastaliq; Hindi, Marathi and Nepali in Devanagari; Amharic and Tigrinya in Ge'ez; Tibetan; Chinese and Japanese; and historical manuscripts through handwritten text recognition (HTR).

We'll publish coverage for each script, marked production, beta or research, so you know what to expect before you send a page.

Until Rune-1 is out, you can try the preview models free on your own pages: Persian, Arabic, Urdu, Hindi, Hebrew, Chinese, Japanese, Korean, Russian and German, Fraktur too.

Run Rune-1 in the cloud, on your servers or on a device.

Use the API, run the model inside your own network, or ship it in an app. The output format is the same in all three.

Cloud API

Send pages to api.runemic.com and get the text back. Nothing to install.

Your servers

Run Rune-1 inside your own network, so documents never leave it.

On a device

Rune-1 Small runs on a laptop or in an app, even offline.

RuneBench: an open OCR benchmark for low-resource scripts

An open benchmark for OCR on low-resource scripts. We report character error rate for each script separately instead of one average. The dataset, the scoring code and every engine's output are public, including the pages where we do badly. We'll also report results on established public benchmarks, so Rune can be compared with other systems.

Public: Persian

RuneBench on GitHub

Test the API while Rune-1 trains

The endpoint Rune-1 will use is live now, with seven preview models. Build against it today; moving to Rune-1 later means changing the model name.

  • Free at runemic.com/try, no sign-up
  • Sign in for 2,500 free credits, no card needed
  • Compare up to three models, and get an API key with cURL, Python and JavaScript examples
# one page in, Markdown out
curl https://api.runemic.com/v1/ocr \
  -H "Authorization: Bearer $RUNEMIC_KEY" \
  -F "file=@page.png"

# → { "text": "# Invoice No. 4821 …",
    "usage": { "credits_used": 2 } }