🤗 Model • 🗂️ Dataset • 📄 Paper

📚 Docs • 🏆 Leaderboard •

Churro is the fastest way to turn hard-to-read historical scans into reliable text. It gives researchers, libraries, archives, and product teams a unified OCR toolkit for handwritten and printed sources, combining high accuracy, low operating cost, and a clean Python API and CLI workflow.

Supported OCR Models and Backends

We provide first-party support for Churro VLM, the best OCR model for historical documents.

Churro also includes built-in profiles, templates, and post-processing for many other models and integrations, including:

• Hosted vision-language models, including Gemini, GPT, Claude, and more, through LiteLLM integration
• OpenAI-compatible servers, including vLLM, Ollama, TGI, and more
• Azure Document Intelligence
• Mistral OCR
• Chandra OCR
• DeepSeek OCR
• Dots OCR
• MinerU
• Infinity Parser
• PaddleOCR VL
• LFM VL

Quick Start

Python 3.12+ and uv are required.

uv tool install churro-ocr
churro-ocr install hf
churro-ocr transcribe --image scan.png --backend hf --model stanford-oval/churro-3B

For more in-depth information, see the Getting Started guide.

Churro Model and Dataset

• Churro 3B VLM exceeds the accuracy of Gemini 2.5 Pro at 15.5x lower cost.
• Churro-DS dataset contains ~100K pages from 155 historical collections spanning 22 centuries and 46 language clusters.

_{Cost vs. accuracy: Churro (3B) achieves higher accuracy than much larger commercial and open-weight VLMs while being substantially cheaper.}

The following are pages from the CHURRO dev set, randomly picked from the subset where Churro outperforms Gemini 2.5 Pro on the main metric, Normalized Levenshtein Similarity (NLS).

Citation

If you use CHURRO or CHURRO-DS, please cite:

@inproceedings{semnani2025churro,
	title        = {{CHURRO}: Making History Readable with an Open-Weight Large Vision-Language Model for High-Accuracy, Low-Cost Historical Text Recognition},
	author       = {Semnani, Sina J. and Zhang, Han and He, Xinyan and Tekg{"u}rler, Merve and Lam, Monica S.},
	booktitle    = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025)},
	year         = {2025}
}

License

• Code: Apache 2.0
• Model weights: Qwen research license
• Dataset: research use only because of the underlying source licenses