Newer Models, Same Advantage

A specialized optical character recognition model designed for Brazilian Portuguese outperformed newer, more broadly capable multilingual models on Portuguese-language documents. The advantage stems from a fundamental design principle: when a model concentrates all its parameters on a single language and domain rather than distributing them across multiple languages, it achieves better performance on that specific task. The model was trained through two stages: supervised fine-tuning on Portuguese-language documents to build domain competency, followed by preference optimization to improve stability and reduce errors in production conditions. This demonstrates that architectural advancement alone does not necessarily overcome the structural advantage gained through domain specialization and targeted training.