datalab-to/chandra

Datalab

State of the Art models for Document Intelligence

Chandra OCR 2

Chandra OCR 2 is a state of the art OCR model that converts images and PDFs into structured HTML/Markdown/JSON while preserving layout information.

Try Chandra on Datalab

Our managed platform runs an improved Chandra with higher accuracy than the open weights, zero data retention by default, SOC 2 Type 2, and custom BAAs.

If you have high volume workloads, we offer a batch processing service that has processed 200M+ pages per week — we manage the infrastructure so your workloads finish on time.

Get started with $5 in free credits — sign up — takes under 30 seconds — or try Chandra in our public playground.

Commercial self-hosting requires a license — see Commercial usage. For on-prem licensing, contact us.

News

• 3/2026 - Chandra 2 is here with significant improvements to math, tables, layout, and multilingual OCR
• 10/2025 - Chandra 1 launched

Features

• Tops external olmocr benchmark and significant improvement in internal multilingual benchmarks
• Convert documents to markdown, html, or json with detailed layout information
• Support for 90+ languages (benchmark below)
• Excellent handwriting support
• Reconstructs forms accurately, including checkboxes
• Strong performance with tables, math, and complex layouts
• Extracts images and diagrams, and adds captions and structured data
• Two inference modes: local (HuggingFace) and remote (vLLM server)

Quickstart

The easiest way to start is with the CLI tools:

pip install chandra-ocr

# With vLLM (recommended, lightweight install)
chandra_vllm
chandra input.pdf ./output

# With HuggingFace (requires torch)
pip install chandra-ocr[hf]
chandra input.pdf ./output --method hf

# Interactive streamlit app
pip install chandra-ocr[app]
chandra_app

Benchmarks

Multilingual performance was a focus for us with Chandra 2. There isn't a good public multilingual OCR benchmark, so we made our own. This tests tables, math, ordering, layout, and text accuracy.

See full scores below. We also have a full 90-language benchmark.

We also benchmarked Chandra 2 with the widely accepted olmocr benchmark:

See full scores below.

Examples

Installation

Package

# Base install (for vLLM backend)
pip install chandra-ocr

# With HuggingFace backend (includes torch, transformers)
pip install chandra-ocr[hf]

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

If you're using the HuggingFace method, we also recommend installing flash attention for better performance.

From Source

git clone https://github.com/datalab-to/chandra.git
cd chandra
uv sync
source .venv/bin/activate

Usage

CLI

Process single files or entire directories:

# Single file, with vllm server (see below for how to launch vllm)
chandra input.pdf ./output --method vllm

# Process all files in a directory with local model
chandra ./documents ./output --method hf

CLI Options:

• --method [hf|vllm]: Inference method (default: vllm)
• --page-range TEXT: Page range for PDFs (e.g., "1-5,7,9-12")
• --max-output-tokens INTEGER: Max tokens per page
• --max-workers INTEGER: Parallel workers for vLLM
• --include-images/--no-images: Extract and save images (default: include)
• --include-headers-footers/--no-headers-footers: Include page headers/footers (default: exclude)
• --batch-size INTEGER: Pages per batch (default: 28 for vllm, 1 for hf)

Output Structure:

Each processed file creates a subdirectory with:

• <filename>.md - Markdown output
• <filename>.html - HTML output
• <filename>_metadata.json - Metadata (page info, token count, etc.)
• Extracted images are saved directly in the output directory

Streamlit Web App

Launch the interactive demo for single-page processing:

chandra_app

vLLM Server (Optional)

For production deployments or batch processing, use the vLLM server:

chandra_vllm

This launches a Docker container with optimized inference settings. Configure via environment variables:

• VLLM_API_BASE: Server URL (default: http://localhost:8000/v1)
• VLLM_MODEL_NAME: Model name for the server (default: chandra)
• VLLM_GPUS: GPU device IDs (default: 0)

You can also start your own vllm server with the datalab-to/chandra-ocr-2 model.

Configuration

Settings can be configured via environment variables or a local.env file:

# Model settings
MODEL_CHECKPOINT=datalab-to/chandra-ocr-2
MAX_OUTPUT_TOKENS=12384

# vLLM settings
VLLM_API_BASE=http://localhost:8000/v1
VLLM_MODEL_NAME=chandra
VLLM_GPUS=0

Commercial usage

This code is Apache 2.0, and our model weights use a modified OpenRAIL-M license (free for research, personal use, and startups under $2M funding/revenue, cannot be used competitively with our API). To remove the OpenRAIL license requirements, or for broader commercial licensing, visit our pricing page here.

Benchmark table

Multilingual benchmark table

The table below covers the 43 most common languages, benchmarked across multiple models. For a comprehensive evaluation across 90 languages (Chandra 2 vs Gemini 2.5 Flash only), see the full 90-language benchmark.

Full 90-language benchmark table

We also have a more comprehensive evaluation covering 90 languages, comparing Chandra 2 against Gemini 2.5 Flash. The average scores are lower than the 43-language table above because this includes many lower-resource languages. Chandra 2 averages 72.7% vs Gemini 2.5 Flash at 60.8%.

See the full 90-language results.

Throughput

Benchmarked with vLLM on a single NVIDIA H100 80GB GPU using a diverse mix of documents (math, tables, scans, multi-column layouts) from the olmOCR benchmark set. This set is significantly slower than real-world usage - we estimate 2 pages/s in real-world usage.

Credits

Thank you to the following open source projects:

• Huggingface Transformers
• VLLM
• olmocr
• Qwen 3.5