Lisbon, Dec. 29, 2025 (Lusa) - The Amália artificial intelligence (AI) model performs better in European Portuguese than other open models, according to the technical report from the investment and development team to which Lusa had access on Monday.
"The results show that AMALIA-DPO [Direct Preference Optimisation] achieves the best performance among fully open models by a considerable margin, even obtaining the best results among all models in lexicology and semantics, demonstrating a robust mastery of the specific linguistic skills" of Portuguese from Portugal in various categories.
The Portuguese LLM [Large Language Model] Amália has been constantly evolving by the consortium of Portuguese universities that leads its development.
According to the technical report, Amália has clear advantages over other open models in an in-depth evaluation of European Portuguese.
In Portuguese national exams (long-answer questions in Portuguese), Amália "scores the highest among all fully open-source models, demonstrating a good understanding of complex statements and coherent text production, with appropriate grammar and register".
In this report, "we present an LLM that prioritises European Portuguese and its cultural context", reads the document, which states that Amália uses data from arquivo.pt and post-training data prepared specifically for European Portuguese.
The document indicates that the LLM was trained with language modelling and instruction tuning strategies.
"A key challenge in developing this model was the lack of benchmarks to monitor the model's performance progress," the report notes.
To mitigate this limitation, "we used PT-PT national exams, created a linguistic benchmark, and translated several series of datasets" with a dedicated high-quality machine translation (MT) model.
"The evaluation showed that Amália outperforms all previous open-source models on PT-PT and many open-weight models [which share weights (trained parameters)]," the technical report concludes.
Experiments on language comprehension and inference benchmarks show state-of-the-art or comparable results, while on language generation benchmarks, the model excels in the quality of the generated text. Safety experiments also show that the model is in line with the state of the art," the report reads.
In the future, "we will explore other reinforcement learning methods and develop new combinations of training data to improve reasoning capabilities in PT-PT."
In other words, in practice, these results indicate that Amália is becoming reliable as an assistant in European Portuguese.
The report was prepared by João Magalhães (UNL) and André Martins (IST), coordinators, and a team of about 20 people from the University of Lisbon and Universidade Nova de Lisboa.
The Amalia model was developed by a team composed of the Universidade Nova de Lisboa, the Instituto Superior Técnico, the University of Coimbra, the University of Porto, the University of Minho, and the Foundation for Science and Technology.
The process of creating Amália began with the large-scale collection and processing of data in European Portuguese, which was filtered for relevance and linguistic quality. For this purpose, the Portuguese Web Archive was used. The model was pre-trained with this data and then fine-tuned on other data to follow instructions, reason, and solve problems.
To train the models, a large-scale computational infrastructure was used, including national supercomputers (Mare Nostrum 5 and Deucalion) and European supercomputers (through the EuroHPC network).
ALU/ADB // ADB.
Lusa
