Chandra OCR : Beats DeepSeek OCR !
How to use Datalab Chandra OCR?
OCR models have been around long enough to make bold claims about “document intelligence,” yet most still break down the moment you throw something slightly chaotic at them: a blurred PDF, an old math worksheet, a scanned form from the 80s.
https://medium.com/media/e9cfb3eda7127bb3e836559d4df03c6a/href
DeepSeek OCR did a fine job on clean pages, but it never quite understood the structure. Then came Chandra OCR, a model that doesn’t just read text but reconstructs the whole idea behind it.
This thing outputs Markdown, HTML, or JSON with precise layout data intact. You don’t just get text, you get a living replica of the original document.
The Core Idea
Chandra was built around a simple but ruthless goal: preserve structure. Most OCR tools flatten everything into a wall of text. Chandra keeps it all, the forms, tables, footnotes, columns, math, and even handwritten scribbles.
It doesn’t matter if it’s a physics problem set or an insurance form, the model reassembles everything the way it was meant to be read. It can output Markdown for simplicity, HTML for rendering, or JSON if you’re pulling it into a data pipeline.
It’s not just recognizing text; it’s reconstructing intent.
Why It Matters
DeepSeek OCR was decent, until you put it up against real-world garbage: multi-column news scans, old textbooks, handwritten notes. Chandra doesn’t fold under that kind of mess. It identifies headers, columns, and page footers with uncanny accuracy. Even handwriting, messy or stylized, gets parsed into coherent sentences.
And then there’s the multilingual edge: Chandra supports 40+ languages out of the box. It doesn’t blink at mixed-language PDFs or math-heavy pages. It even extracts embedded images and diagrams with captions and tags, so you get a document that actually remembers what it looked like.
Under the Hood
The model sits on top of VLLM and Hugging Face infrastructure, optimized for both speed and scale. Whether you’re running it locally or on a cluster, it’s flexible enough to handle batch inputs or interactive exploration through their Streamlit app.
The output can be structured into Markdown, HTML, or JSON, each retaining line breaks, spacing, and visual hierarchy. You can even choose between vLLM for fast server-side inference or Hugging Face if you want direct access to the transformer weights.
It’s made by Datalab, and like their other projects, it integrates cleanly with open-source frameworks. Nothing proprietary, no black-box behavior.
Benchmarks: Where It Stands
Numbers don’t tell the whole story, but they matter here. Using the olmocr benchmark, which is basically the OCR equivalent of ImageNet, Chandra crushed everything else.
It scored 83.1 ± 0.9 overall, outperforming DeepSeek OCR’s 75.4 ± 1.0, dots.ocr’s 79.1, and olmOCR’s 78.5. Even GPT-4o and Gemini Flash 2 couldn’t keep up, especially on multi-column and old-scan tests.
In more practical terms: Chandra doesn’t lose its mind when faced with low-quality scans, small fonts, or math diagrams. It reconstructs forms accurately, even checkboxes. You could digitize an entire legal contract or a school exam set and get a clean, editable structure out of it.
Real-World Examples
The examples tell their own story. A doctor’s handwritten note becomes structured Markdown with sections and line breaks. A geography textbook scan comes back as readable, layout-preserved text with images and captions in place. Even the New York Times archive pages, nightmares for most OCRs, are split perfectly into multi-column Markdown.
There’s even support for math-heavy material. Worksheets, formulas, diagrams, everything holds its relative structure, not just the content. It’s the kind of fidelity that makes it actually usable for downstream LLM tasks like summarization or document Q&A.
Developer Friendly
It’s easy to start with, install, run, and you’re done. You can choose between the CLI, the hosted playground, or a simple interactive Streamlit interface. The output folder gives you neat, structured text files ready for any workflow, RAG pipelines, document indexing, or search retrieval.
Verdict
Chandra OCR doesn’t just edge past DeepSeek, it resets what good OCR looks like. It reads like a human and writes like a machine that actually respects structure.
It’s one of those rare tools that feels designed for the post-LLM world, where text extraction isn’t enough, and structure is the story. If you’re building document pipelines or AI reading systems, you can stop stitching together half-broken OCRs.
Chandra finally reads the page the way it was meant to be read.
datalab-to/chandra · Hugging Face
Chandra OCR : Beats DeepSeek OCR ! was originally published in Data Science in Your Pocket on Medium, where people are continuing the conversation by highlighting and responding to this story.
