stanford-oval

CHURRO: Making History Readable with an Open-Weight Large Vision-Language Model for High-Accuracy, Low-Cost Historical Text Recognition Handwritten and printed text recognition across 22 centuries and 46 language clusters, including historical and dead languages. Cost vs. accuracy: CHURRO (3B) achieves higher accuracy than much larger commercial and open-weight VLMs while being substantially cheaper. CHURRO is a 3B-parameter open-weight vision-language model (VLM) for historical document transcription. It is trained on CHURRO-DS, a curated dataset of ~100K pages from 155 historical collections spanning 22 centuries and 46 language clusters. On the CHURRO-DS test set, CHURRO delivers 15.5× lower cost than Gemini 2.5 Pro while exceeding its accuracy. For more details and code see https://github.com/stanford-oval/Churro.