Portugal's Sovereign AI Model Amalia Nears Final Release — 60 Researchers Built an LLM That Actually Speaks European Portuguese

Portugal is building its own large language model — and it is named after the country's most iconic cultural figure. Amalia, a sovereign AI system purpose-built for European Portuguese, is on track for its final release in June 2026 after a consortium of seven universities and more than 60 researchers spent 19 months training it on national supercomputers.

Why Portugal Needs Its Own AI

When a Portuguese citizen asks ChatGPT, Gemini, or Claude a question in Portuguese, the response almost invariably comes back in Brazilian Portuguese — a linguistically and culturally distinct variant. European Portuguese (pt-PT) has different grammar, vocabulary, spelling, and cultural references, yet it is dramatically underrepresented in the training data of major commercial AI models.

This is not just a matter of national pride. When AI systems are deployed in healthcare, education, or public administration, linguistic accuracy matters. A medical chatbot that uses Brazilian terminology for Portuguese patients, or an educational tool that teaches Brazilian spelling conventions to Portuguese children, creates real problems.

Amalia — formally the Multimodal Automatic Language Assistant with Artificial Intelligence, named in homage to fado legend Amália Rodrigues — is designed to close that gap.

Who Built It

The project is coordinated by the Centre for Responsible AI, led by Paulo Dimas (also vice-president of innovation at Unbabel), and involves a consortium of Portuguese universities:

• NOVA School of Science and Technology (lead technical team, coordinated by Prof. João Magalhães)
• Instituto Superior Técnico (co-coordinated by Prof. André Martins)
• Universidade de Coimbra
• Universidade do Porto
• Universidade do Minho
• Universidade da Beira Interior
• Universidade de Évora

The model was trained on MareNostrum 5 in Barcelona (where Portugal holds computing capacity), the Portuguese national supercomputer Deucalion, and EuroHPC network resources — keeping costs far below what commercial cloud training would require.

Training data included material from arquivo.pt, Portugal's national web archive, with post-training specifically prepared for European Portuguese cultural and linguistic norms.

How It Performs

According to a technical report published on arxiv in March 2026, Amalia is the best-performing fully open-source model for European Portuguese by a considerable margin. Key results:

• The Amalia-DPO variant (trained with Direct Preference Optimisation) achieves the best results of all models — open and open-weight — in lexicology and semantics
• On Portuguese national exam questions (long-answer format), Amalia scores highest among all fully open models, demonstrating comprehension of complex sentences and coherent text production
• On translated standard benchmarks, it matches strong international baselines
• On new pt-PT-specific evaluations created by the team, it substantially outperforms all other models

The team also released a new suite of European Portuguese benchmarks, including tasks targeting generation quality, linguistic competence, and the ability to distinguish between pt-PT and pt-BR — filling a gap in AI evaluation infrastructure.

Who Can Use It

The beta version is already available to the Portuguese Public Administration through the FCT's IAedu platform. When the final version launches in June, it will be released as a fully open-source model — free for public entities, businesses, startups, researchers, and citizens.

Planned use cases include:

• Healthcare: The Halo project (by Unbabel) is using Amalia to restore communication for ALS patients through AI voice cloning in proper European Portuguese
• Education: A personalised AI tutor aligned with Portugal's national curriculum
• Public services: Integration across government digital platforms via AMA and the gov.pt portal
• Startups: A foundation model for Portuguese-language AI applications

The Bigger Picture

Amalia sits within a broader European movement toward digital sovereignty — the idea that countries and regions should not be entirely dependent on American or Chinese technology companies for critical AI infrastructure. France has Mistral, Germany is investing in sovereign cloud capacity, and now Portugal has a model that ensures its language and culture are represented in the AI age.

Prime Minister Luís Montenegro announced the project at Web Summit in November 2024, calling it part of Portugal's National Artificial Intelligence Agenda. With a final release expected in two months, the question now is whether Portuguese public institutions and businesses will adopt it at scale — or whether commercial models will continue to dominate despite their Brazilian Portuguese default.