Portugal releases Amália, the first national AI model for European Portuguese

In July 2026, Portugal released Amália, its first national language model, built specifically for European Portuguese. The country, with a population of just over 10 million, chose the most transparent route possible: it published not only the model weights but also the training data the model is built on.

Why European Portuguese needs its own model

To English or Spanish speakers, the difference between European and Brazilian Portuguese may seem minor. For users and developers, it is fundamental. This is not just a matter of accent: the two varieties have different lexical norms, syntactic constructions, idioms and spelling rules, the latter fixed after the 2009 reform.

Brazil, with its 200 million people, generates vastly more digital content than Portugal. Large international language models are trained on whatever is most abundant online, and European Portuguese loses out to the Brazilian variety by a wide margin. For Portuguese users, the effect is concrete: models mix up forms of address, mishandle official administrative language, and poorly understand local idioms and legal constructions. Amália was developed specifically to close this gap.

What the government has made public

Key facts about the project:

• Amália is Portugal's first official national language AI model
• It is trained specifically on European Portuguese, not Brazilian
• Open access: the model weights and the training dataset are published
• The initiator is the Portuguese state; the country's population is about 10 million
• Release date: July 2026

Publishing training data is rare even among academic open projects. It lets independent researchers and developers do more than use a finished model: they can see exactly what it was trained on, fine-tune it for specific tasks, and check its quality without a black box. Government agencies, universities, startups and medical institutions get access without licensing barriers.

Amália and European AI sovereignty

The release fits into a Europe-wide debate about how smaller countries can avoid depending entirely on American and Chinese language models. The EU AI Act, adopted in 2024, encourages member states to build their own technological capabilities, and Portugal is applying that principle literally.

For small and mid-sized European languages such as Dutch, Czech, Greek, Hungarian and Finnish, the gap in AI processing quality compared with English is especially visible. Large labs have little economic incentive to invest proportionate resources in languages with 10–50 million speakers: the market does not produce quality models where there is not enough training data or commercial appeal. State initiatives like Amália fill exactly this niche, and an open license turns the model into a public good that the whole community can build on and improve.

What this means

Portugal has set a working precedent: a country of 10 million people can build and open up a language model for its own variety of a language. If the approach proves reproducible, and open data and weights allow others to repeat it, it could become a template for the Netherlands, Czechia, Greece and other European states whose languages remain on the periphery of global AI systems.