Chandra OCR 2 reads handwriting in 90 languages — free

Scanning a document sounds simple — until it's a handwritten form, a table full of numbers, or a page of math equations. Most AI scanning tools either jumble the layout or skip the hard parts entirely. Chandra OCR 2 just changed that.

Released by Datalab, this open-source OCR model (OCR stands for Optical Character Recognition — the technology that converts images of text into actual text you can copy, search, and edit) scored 85.9% on the standard OCR benchmark, beating every alternative tested. And it works across 90 languages — from English to Arabic, Japanese, Hindi, and Russian.

It beats GPT-5 Mini and Gemini at reading documents

The numbers tell the story. On a 43-language test, Chandra 2 scored 77.8% average accuracy. Google's Gemini 2.5 Flash scored 67.6%. OpenAI's GPT-5 Mini scored 60.5%. That's not a marginal lead — Chandra beats the big names by 10 to 17 percentage points.

On the expanded 90-language benchmark, the gap stayed wide: Chandra at 72.7% vs. Gemini at 60.8%.

Not just text — it preserves the layout

Most OCR tools strip formatting. Chandra 2 keeps it. Multi-column layouts, headers, footers, table borders, nested cells — all preserved in the output. You get clean Markdown, HTML, or JSON that mirrors the original document's structure.

This matters for anyone who processes invoices, legal contracts, academic papers, or government forms. Instead of manually fixing garbled output, you get something you can immediately use.

Accountants, researchers, and admins — this is for you

If you process invoices or receipts, Chandra can extract structured data from scanned documents in seconds. If you work with academic papers, it handles math equations and scientific notation that other tools mangle. If you handle multilingual documents, the 90-language support means you don't need separate tools for each language.

The model processes about 2 pages per second on a modern GPU, or you can use Datalab's hosted API with a free tier for smaller workloads.

Try it now

Install locally in one line:

pip install chandra-ocr[all]

Then scan a document:

chandra input.pdf ./output --method hf

Or launch the visual web interface:

chandra_app

No GPU? Try the free online playground to test it instantly in your browser. The code is open-source on GitHub under Apache 2.0, free for personal use and startups under $2M revenue.

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News