Portugal has officially launched Amália, its first national open-source large language model, releasing the model, its datasets, and source code under an open-source license so that governments, businesses, universities, and researchers can reuse and build on it. Unveiled at the start of July 2026, Amália is designed specifically for European Portuguese and is intended not as a consumer chatbot but as public digital infrastructure that organizations can adapt for their own use cases.

The launch is the latest move in Europe's broader push for sovereign AI, and it lands amid a fast-moving stretch of breaking AI news in which nations and blocs are racing to build models they can control. Where France continues to debate how to anchor an independent AI strategy, Portugal has quietly delivered one — reportedly for roughly €7 million in total, with an initial €5.5 million drawn from the country's Recovery and Resilience Plan (PRR).

What Amália actually is

Amália is a Portuguese-tuned language model built to understand and generate the European variant of Portuguese, which is often underrepresented in models trained predominantly on English and Brazilian Portuguese data. Rather than training a foundation model entirely from scratch, the project builds on EuroLLM-9B, the European open foundation model. Researchers then expanded it with European Portuguese datasets, increased its model capacity and context window, enhanced its safety and evaluation systems, and added multimodal text-and-image capabilities.

The result is a model that the Portuguese government frames as a shared national resource. Because the weights, training data, and code are open, public bodies and companies can fine-tune Amália for their own domains without depending on a single foreign vendor or sending sensitive data abroad.

Who built it, and where it runs

More than 60 researchers and students from Portuguese universities and research institutions contributed to the project. Major allocations from the PRR funding went to NOVA University Lisbon, Instituto Superior Técnico, and the Universities of Porto, Minho, and Coimbra, along with the Foundation for Science and Technology (FCT).

Training and experimentation leveraged serious compute: the project used the Deucalion supercomputer, based in Portugal, and Spain's MareNostrum 5 machine. A test version of Amália was completed in September 2025 and presented at the PROPOR conference in Brazil, an international meeting focused on the computational processing of Portuguese. Funding has reportedly already been secured for the project's next phase, which runs through 2027.

Where it will be deployed first

The first deployments target Portugal's public sector, spanning education, defense, healthcare, culture, and citizen services. Planned applications include:

  • An AI teaching assistant for classrooms
  • A virtual guide for Portuguese museums and monuments
  • A digital citizen services assistant for public administration
  • Decision-support tools for the Portuguese Navy

By focusing on these use cases, Portugal is effectively treating Amália as the kind of shared platform that other countries have tried — with mixed success — to build through national AI strategies. The open-source approach means the work can be audited, extended, and reused rather than locked inside a proprietary API.

Why sovereign AI keeps gaining ground

Amália arrives at a moment when the debate over AI sovereignty has moved from rhetoric to deliverables. European institutions have warned for years that relying on a handful of U.S. labs creates strategic and cultural risks, from language coverage to data residency. Portugal's bet is that a relatively small, focused investment — a fraction of what frontier labs spend in a week — can produce a genuinely useful national model when it builds on open foundations and targets a specific language.

The model is also a data-sovereignty play. For sensitive public-sector work, such as defense decision-support or healthcare triage tools, the ability to run a model on national infrastructure, inspect its training data, and fine-tune it locally is a feature that closed commercial APIs cannot easily match.

The road ahead

The open question is adoption. An open model is only as valuable as the ecosystem that grows around it, and Portugal will need developers, integrators, and public agencies to actually build on Amália rather than default to familiar commercial tools. The team's decision to release datasets and evaluation systems alongside the weights is aimed squarely at lowering that barrier.

For now, Amália stands as one of the clearest examples of a smaller European country shipping a concrete sovereign-AI product rather than just commissioning a study. Whether it becomes the backbone of Portuguese public AI — or a proof of concept that larger nations borrow from — will depend on what gets built on top of it next.

Stay Ahead of AI

Portugal's Amália is just the latest signal that the global AI race now runs through national and open-source models, not only the familiar frontier labs. For ongoing coverage of model releases, policy shifts, and the business of artificial intelligence, follow the latest AI developments at AI Buzz Wire.

Read more AI news →