---
type: "post"
id: "release-indra-ocr-1-1-0-datalab-to-ac3594"
slug: "release-indra-ocr-1-1-0-datalab-to-ac3594"
title: "Release Indra OCR 1.1.0"
description: "Datalab’s Indra OCR 1.1.0 adds handwriting recognition, more than 80 languages on scans, and correction for rotated pages while remaining CPU-runnable."
retrieval_nugget: "Datalab’s Indra OCR 1.1.0 adds handwriting recognition, more than 80 languages on scans, and correction for rotated pages while remaining CPU-runnable. Its release reports improvements on OmniParseBench, especially for handwriting and rotated pages, alongside a larger default model download. The update matters to teams parsing scanned documents locally."
published_at: "2026-11-11"
updated_at: "2026-11-11"
record_date: "2026-11-11"
date_kind: "published_at"
topics: ["multimodal"]
entities: ["Datalab","Indra OCR","OmniParseBench"]
related_patterns: []
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://github.com/datalab-to/indra/releases/tag/v1.1.0"
source_urls: ["https://github.com/datalab-to/indra/releases/tag/v1.1.0"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:release-indra-ocr-1-1-0-datalab-to-ac3594"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/release-indra-ocr-1-1-0-datalab-to-ac3594/","markdown":"https://newruntime.com/posts/release-indra-ocr-1-1-0-datalab-to-ac3594.md","json":"https://newruntime.com/posts/release-indra-ocr-1-1-0-datalab-to-ac3594.json"}
markdown_url: "https://newruntime.com/posts/release-indra-ocr-1-1-0-datalab-to-ac3594.md"
---

# Release Indra OCR 1.1.0

## Retrieval answer

Datalab’s Indra OCR 1.1.0 adds handwriting recognition, more than 80 languages on scans, and correction for rotated pages while remaining CPU-runnable. Its release reports improvements on OmniParseBench, especially for handwriting and rotated pages, alongside a larger default model download. The update matters to teams parsing scanned documents locally.

## Summary
Datalab’s Indra OCR 1.1.0 adds handwriting recognition, more than 80 languages on scans, and correction for rotated pages while remaining CPU-runnable. Its release reports improvements on OmniParseBench, especially for handwriting and rotated pages, alongside a larger default model download. The update matters to teams parsing scanned documents locally.

## What happened
Datalab’s GitHub release for Indra OCR 1.1.0 was published on 19 October. The release describes Indra as an OCR system that can run on CPU without a GPU. It adds a line detector and Latin recognizer trained on real handwritten and filled-form pages. For scans, the release says the software now supports more than 80 languages, with recognizers for Chinese, Japanese and Korean; Arabic, Persian, Urdu, Pashto and Hebrew; Indic scripts; Thai, Lao, Khmer, Myanmar, Sinhala, Georgian, Armenian and Ethiopic. Each line is routed to a recognizer for its script, and language models re-read lines the system considers uncertain.

The release also adds page-orientation handling. It says Indra straightens tilts of up to 45 degrees, turns sideways and upside-down scans upright, and can read sideways tables in a PDF’s text layer correctly. Other listed changes include recognizing light text on dark backgrounds, such as slides and dark table headers, and no longer mistaking Chinese text blocks for tables. The repository includes an image-to-Markdown, text or JSON example and a quickstart notebook covering image files, in-memory images and math.

Datalab reports results from the OmniParseBench leaderboard using the benchmark’s Indra adapter and scorer. Comparing Indra OCR 0.3.2 with 1.1.0, the headline score rises from 62.3 to 66.7, or from 67.0 to 71.4 with `--math`. The release lists handwriting tests at 8.3% and 25.2%, rotated pages at 25.3% and 50.3%, and tables at 56.3 and 61.1, respectively. It also reports approximately 44 pages per second on 32 x86 CPU cores using an AMD EPYC system, described as 15 times Tesseract and 10 times PaddleOCR’s PP-OCR text models on the same machine type and pages; on four cores, it reports about 210 milliseconds per page.

The default model download is now about 76 MB, up from 18 MB: the release attributes 30 MB to networks and 46 MB to language models. Users of 0.3.x retain their pinned models, listed as v3.1 and v3.1.1, until they upgrade. Installation is `pip install -U indra-ocr`; the release points to the `datalab-to/indra-ocr` Hugging Face models tagged v1.1.0. It also notes that `--max-pages` now rejects negative values.

## What the archive adds
New Runtime’s earlier coverage on 17 October described Indra OCR as a compact Datalab OCR model, around 18 MB, running through ONNX Runtime without PyTorch. Version 1.1.0 is a new event in that ongoing product story: it adds recognition and orientation capabilities, with the default download now around 76 MB. An earlier note on [Unlimited-OCR 2](https://newruntime.com/posts/independent-16-gb-gpu-test-finds-unlimited-ocr-33a058/) examined document-parsing performance under memory constraints, but it is a separate system rather than a direct test of Indra.

## Trend signal
This does not yet fit a pattern New Runtime tracks. The release is a concrete OCR update, but it does not by itself establish a broader shift in agent architecture, economics or deployment practice.

## Why now
The release supplies a specific upgrade point: 0.3.x users keep their pinned model versions until they choose to upgrade. Teams can therefore test the larger default download and the new script and orientation handling against their own scan mix before changing production deployments.

## Why it matters
For teams building document pipelines, handwriting, multilingual scans and misoriented pages are common sources of manual cleanup or routing to separate tools. Combining these capabilities in a CPU-runnable package could reduce that branching for some workloads, particularly where GPU deployment is unavailable or unjustified. The tradeoff is an increase in the default download from 18 MB to about 76 MB, and benchmark gains are not a guarantee for every language, document layout or scan quality.

The useful engineering question is not simply whether the leaderboard score rose; it is whether the upgrade reduces correction work on the documents the team actually receives. Test representative handwritten forms, scripts and rotated pages, then compare field-level accuracy, processing time and review burden with the current pipeline. Keep a fallback for cases where specialized OCR or human verification remains necessary.

## Limits and uncertainty
The benchmark figures are reported by Datalab using OmniParseBench’s Indra adapter and scorer; the release does not establish independent replication. Its speed comparisons apply to the stated CPU setup and pages, not every machine or document mix. The reported gains do not specify performance for each of the more than 80 languages, handwriting style, image quality or layout. The release points readers to the README for limitations, and production accuracy, memory use and review savings will depend on local data. The larger download is explicit, but the release does not quantify any other deployment costs.

## Skill to build now
Build a small OCR regression set from representative documents already permitted for testing: include handwritten forms, each script your pipeline encounters, rotated pages, dark backgrounds and tables. Run the current version and 1.1.0 on the same inputs, and compare extracted fields against reviewed ground truth as well as latency and manual corrections. Keep the set as an upgrade gate so a broader benchmark score does not substitute for your own acceptance criteria.

## Reader takeaway
Indra OCR 1.1.0 broadens what a compact CPU OCR pipeline can attempt, particularly handwriting, multiple scripts and rotated scans. Its reported benchmark and speed improvements are promising, but the larger model download and dataset-specific accuracy mean teams should validate on their own documents before upgrading.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)