AMALIA-VL: A Native European Portuguese Open-Source Vision and Language Model

A team of researchers has introduced AMALIA-VL, the first open-source instruction-tuned Large Vision and Language Model developed natively for European Portuguese, addressing a long-standing gap in multilingual AI resources [1]. The model, detailed in a paper submitted on 17 June 2026, is designed to serve the specific linguistic needs of European Portuguese (pt-PT), a variant the authors say is systematically underserved by existing open-source multimodal models. These models, they note, either conflate pt-PT with Brazilian Portuguese or severely under-represent it in their training data mixes [1][2]. AMALIA-VL pairs a high-resolution vision encoder with dynamic image tiling and a fully open pt-PT-optimized language model via a learned connector [1]. The development process involved a three-stage training regimen: vision-language alignment, general visual instruction tuning, and preference optimization. This was supported by a pt-PT-centric multimodal data mix that combines curated and translated public datasets with novel datasets created to address what the researchers describe as the near-total absence of European Portuguese multimodal resources [1][2]. The project was led by Diogo Glória-Silva [1]. To foster further development, the team has committed to releasing the model weights, training data, and construction pipelines. They will also provide machine-translated pt-PT evaluation benchmarks, a move intended to democratize the development of European Portuguese Large Vision and Language Models [1][2]. The submission, weighing 3,562 KB, establishes a strong baseline for future open-source pt-PT models, according to the researchers' evaluation [1]. The work arrives as the broader field of Large Vision and Language Models continues to advance rapidly, yet linguistic diversity in training data remains a persistent challenge. The release of a fully open model for a lower-resource language variant like European Portuguese provides a new tool for researchers and developers working on multimodal AI applications in the region [2].