AMALIA is the first Large-Scale Language Model (LLM) designed specifically for European Portuguese, using AI to process and generate natural language, with the ability to understand the linguistic diversity, cultural context, and historical references unique to Portugal.
AMALIA is a multimodal model, capable of interpreting text, speech, images and video, and was developed by national research institutions under the coordination of the State Technological Reform Agency (ARTE).
All versions of the model, along with the training data, will be released as open source, making them freely available to researchers, public institutions, companies, and citizens.
Goals
The main objective of AMALIA is to promote digital sovereignty, ensuring that sensitive national data can be processed securely and independently. It also aims to strengthen the national AI ecosystem by supporting innovation in key sectors such as education, culture, speech, media, science and public administration:
- Promotion of the Portuguese language: Enabling Companies, Public Administration (PA) and other organizations to communicate accurately and effectively in European Portuguese.
- Ensure data sovereignty for citizens: Enabling the application of the latest advances in AI to domains such as PA, while ensuring that sensitive data remains within Portugal’s national territory.
- Preservation of Portugal’s cultural identity: Capturing and preserving proverbs, idiomatic expressions, cultural references, and knowledge relating to national figures and landmarks.
- Promotion of AI research and innovation: Fostering the development of a collaborative R&D ecosystem and contributing to the advancement of Portugal’s expertise at the forefront of AI.
Main features and innovations
AMALIA shows strong performance in European Portuguese, outperforming other open models in lexicology and semantics. The model was trained using large-scale computational infrastructure, including national supercomputers (Mare Nostrum 5, Deucalion) and European infrastructure (EuroHPC), which are also projects funded by the Portuguese government. Technical details include:
- Language model with 9 billion parameters (9B);
- Developed from EuroLLM-9B, a European multilingual model;
- Maximum sequence length extended from 4K to 32K tokens, with the application of RoPEscaling (θ from 10,000 to 1,000,000).
The final, multimodal version was made available in 01JUL26 as open source (https://huggingface.co/amalia-llm), reinforcing national digital sovereignty and promoting full access by any person or entity.
Impact
The impact of AMALIA encompasses technological sovereignty, the enhancement of European Portuguese, and a broad modernization of the Portuguese society in areas such as education (providing support for students and teachers), culture (information about national heritage), science (optimization of literature searches) and public administration (development of user-ready solutions for widespread deployment, with ongoing work for integrating the Portuguese LLM with the virtual assistant of the Gov.pt single public services portal.
