Sina Semnani
@sina_semnani
There are many new VLMs/OCR models out there, but most still fail on historical documents.
We’re releasing CHURRO, a 3B open-weight vision-language model that outperforms Gemini 2.5 Pro on historical OCR while being 15.5× more cost-effective.
From our EMNLP 2025 (Main) paper:
We’re releasing CHURRO, a 3B open-weight vision-language model that outperforms Gemini 2.5 Pro on historical OCR while being 15.5× more cost-effective.
From our EMNLP 2025 (Main) paper:
2 25