New Models, Same Advantage: DharmaOCR Specialization Beats Newer Architectures in Portuguese

Researchers who created the DharmaOCR optical character recognition model for Brazilian Portuguese compared it with the newer Mistral OCR4 and Unlimited-OCR models. Although both latter models represent technically more advanced architectures with novel training methods and datasets, DharmaOCR scored 0.925 against 0.798 for Mistral OCR4 and 0.7587 for Unlimited-OCR in a benchmark targeting exclusively Portuguese. The advantage is attributed to specialization: DharmaOCR was trained only on Brazilian Portuguese, concentrating all its resources on this language domain, while the newer models are multilingual and must distribute their capabilities across many languages. Additionally, the Direct Preference Optimization step in DharmaOCR's training pipeline reduced text degeneration, where the model starts generating incoherent repetitive output on blurry images. Multilingual models often make characteristic errors specifically on Portuguese vocabulary and proper nouns; for example, Mistral OCR4 turned the name of famous Brazilian musician Chico Buarque into "Chico Barque", and Unlimited-OCR into "chico bique". According to the authors, specialization remains a structural advantage that will not disappear with the advent of newer architectures as long as training resources are finite.