This free tool just out-read Gemini 2.5 Flash at documents

A Free Tool Just Beat Google's Best at Reading Documents

When people talk about AI document reading, they usually assume you need to pay for a big-name service. Chandra OCR 2, a free open-source (free software anyone can view, use, and modify) tool released in March 2026, just changed that assumption — and the numbers back it up.

Built by a company called Datalab, Chandra OCR 2 scored 85.9% on the industry-standard olmOCR benchmark (a standardized test used to measure how accurately an AI can read and reproduce text from scanned documents and images). For context, Google's Gemini 2.5 Flash scored 63.8% on the same test. OpenAI's GPT-4o scored 69.9%. Chandra outperformed both by a wide margin.

The tool already has over 7,100 stars on GitHub (a measure of how many developers have bookmarked and endorsed the project), and it's climbing the trending charts fast.

What Chandra OCR 2 Actually Does

OCR stands for Optical Character Recognition — technology that converts photos or scans of text into editable, searchable digital text. You might use it to scan a paper invoice and get a spreadsheet, or to pull text out of a photo of a whiteboard. Chandra OCR 2 takes that concept much further.

The output isn't just plain text. Chandra OCR 2 can convert a document image into structured Markdown, HTML, or JSON (three different formats for organizing digital text and data), while preserving the original layout. If a document has a table with 5 columns and 20 rows, you get back a proper digital table — not a jumbled mess of numbers.

The Numbers: How Far Ahead Is Chandra?

The benchmark (standardized accuracy test) results tell a striking story. Here's how the main olmOCR benchmark breaks down by document type:

Document Type	Chandra 2 Score
Math (ArXiv papers)	90.2%
Tables	89.9%
Headers and Footers	92.5%
Multi-column layouts	83.5%
Small fine print	92.1%
Overall Score	85.9%

For comparison, the closest paid competitors scored: GPT-4o at 69.9% and Gemini Flash 2 at 63.8%. Even Datalab's own paid API service only scored 86.7% — just 0.8 points ahead of the free version.

The multilingual (multiple languages) performance gap is even larger. Across 43 commonly tested languages, Chandra 2 averaged 77.8% while Gemini 2.5 Flash averaged 67.6% — a 10 percentage point lead. Expand to the full 90-language test, and Chandra averages 72.7% versus Gemini 2.5 Flash's 60.8%.

A Smaller Model That Somehow Does More

Here's something unusual: Chandra OCR 2 is actually smaller than its predecessor. The first version of Chandra had 9 billion parameters (a rough measure of an AI model's size and computational complexity). Chandra 2 has 5 billion parameters — a nearly 45% reduction in size — yet beats the older version across every single benchmark category.

This matters for practical use because smaller models run faster and require less computing power. Datalab reports that on a single NVIDIA H100 GPU (a professional graphics card commonly used for AI processing), Chandra 2 processes roughly 1.44 pages per second — which in real-world conditions they estimate reaches 2 pages per second. For a 100-page document, that's under a minute.

The improvement in South Asian languages is particularly dramatic. Compared to Chandra 1, the new version scores: Bengali +27.2 points, Kannada +42.6 points, Malayalam +46.2 points, Tamil +26.9 points, and Telugu +39.1 points. These are languages that most Western OCR tools handle poorly.

How to Install and Run Chandra OCR 2

Getting started requires Python 3.10 or higher (a programming language commonly used for AI projects) installed on your computer. The simplest install command is:

# Basic installation
pip install chandra-ocr

# With HuggingFace support (for running the model locally)
pip install chandra-ocr[hf]

# With all features
pip install chandra-ocr[all]

Once installed, you can use it directly from the command line (the text-based interface on your computer):

# Convert a single PDF to a folder of output files
chandra input.pdf ./output

# Convert all PDFs in a folder
chandra ./documents/ ./output

If you'd rather not use the command line at all, there's also a web-based visual interface. Type chandra_app after installing, and it opens a visual app in your browser where you can drag and drop documents and see the results immediately.

For developers who want to use it inside their own applications, there's also a simple Python code snippet:

from transformers import AutoModelForImageTextToText, AutoProcessor
from chandra.model.hf import generate_hf
from chandra.model.schema import BatchInputItem
from PIL import Image
import torch

# Load the model
model = AutoModelForImageTextToText.from_pretrained(
    "datalab-to/chandra-ocr-2",
    dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()
model.processor = AutoProcessor.from_pretrained("datalab-to/chandra-ocr-2")

# Process an image
batch = [BatchInputItem(image=Image.open("document.png"), prompt_type="ocr_layout")]
result = generate_hf(batch, model)[0]
print(result.markdown)

Who Can Use It for Free — and Who Can't

The code itself is released under the Apache 2.0 license (a widely-used open-source license that allows broad commercial and personal use). The actual AI model (the trained intelligence behind the tool) uses a modified OpenRAIL-M license (a license designed specifically for AI tools that allows free use but places some restrictions).

In practice, this means:

• Free for: Personal projects, academic research, and startups earning or funded under $2 million
• Restricted for: Companies that would use it to directly compete with Datalab's own paid OCR service
• Commercial licenses: Available at datalab.to/pricing for larger businesses

For the vast majority of users — freelancers, small businesses, researchers, students, and individual professionals — the tool is completely free.

Real-World Uses for Non-Technical People

You don't need to be a developer to benefit from what Chandra OCR 2 can do. Here are some practical examples of what this kind of tool enables:

• Marketers — scan competitor print ads, brochures, or old catalogues and get all the text in an editable format instantly
• Office workers — convert stacks of paper invoices, receipts, or contracts into searchable digital files without manual retyping
• Designers — extract text from wireframe photos, whiteboard sessions, or printed mockups
• Researchers and students — digitize handwritten notes, old books, or archival documents in dozens of languages
• Developers — build pipelines (automated sequences of tasks) that process incoming documents automatically

The GitHub repository (the online storage location where the project's code lives) is available at github.com/datalab-to/chandra and the model is also available for download at HuggingFace (a popular platform for sharing AI tools).

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