A country of just over 10 million people now has a large language model that speaks its own version of the language. Portugal has released Amália, its first national AI model built specifically for European Portuguese, and it has done so in about the most deliberate way a government can.
The model, its training data and its source code are all open, free for governments, universities and companies to take and build on.
The name is not incidental. Amália is an acronym, for Automatic Multimodal Language Assistant with Artificial Intelligence, but it also reaches for Amália Rodrigues, the fado singer whose voice is bound up with Portuguese identity.
The model itself is built on EuroLLM-9B, a European foundation model, which a team of more than 60 researchers and students expanded with European Portuguese datasets, a larger context window, stronger safety and evaluation systems, and the ability to handle images alongside text.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
What Amália is not is a rival to ChatGPT in the way most people would encounter one. It will not ship as a consumer chat app. It is designed instead to sit underneath other things, as the layer that other software calls on.
The planned uses include an AI teaching assistant, a virtual guide for Portuguese museums and monuments, a digital assistant for citizen services, and decision-support tools for the Portuguese Navy.
That distinction explains a good deal about why the state is giving the model away rather than charging for access to it.
The project has drawn an initial €5.5m through Portugal’s Recovery and Resilience Plan, with money flowing to NOVA University Lisbon, Instituto Superior Técnico, and the universities of Porto, Minho and Coimbra, coordinated with the Foundation for Science and Technology.
A test version was finished in September 2025 and presented at the PROPOR conference in Brazil. Funding has already been secured through the end of 2027, which reads less like a launch and more like a commitment to keep the thing alive.
Open, in this case, means open. Where the large commercial systems are closed boxes accessed through an interface and a bill, Amália ships with its weights, its datasets and its code published under an open licence, so that anyone can inspect how it was trained, adapt it and run it on their own hardware. That choice is partly ideological and partly practical.
A model a government intends to wire into citizen services and naval decision-support is one it wants to be able to audit, not just trust, and open publication is the surest way to keep that option.
The release lands squarely inside Europe’s wider unease about depending on American and Chinese systems for something as foundational as language.
It follows the OpenEuroLLM alliance, the cross-border effort to train open models on the continent’s own languages, and a run of infrastructure bets that includes Nscale’s €695m data-centre push in Portugal with Microsoft. Whether any of it amounts to genuine independence is contested.
Renting GPUs by the hour, TNW has argued, can produce the illusion of sovereignty rather than the substance of it.
Amália’s strongest card is specificity. European Portuguese is not Brazilian Portuguese, and the big commercial models, trained overwhelmingly on the latter, tend to flatten the difference.
A system that gets the grammar, the idiom and the cultural references right is useful in a way that a larger, blurrier one is not, and that gap matters most for public services expected to speak to citizens in their own register rather than an approximation of it.
The harder question is adoption. Publishing a model openly is one thing. Getting universities, companies and government departments to actually build on it is another, and that second step is where most sovereign-AI ambitions quietly run out of road.
Portugal has funded Amália through 2027 and named the institutions meant to carry it forward. The next two years will show whether it becomes real infrastructure or a well-documented research project with a beautiful name.
