On July 1, 2026, the Portuguese government officially unveiled Amália, which its creators describe as the first open large language model (LLM) developed in European Portuguese. The event, held at the innovation center of Instituto Superior Técnico in Lisbon, marked the open-source release of a model whose base version had been delivered in September 2025. Funded through the Recovery and Resilience Plan (PRR), with public investment rising to €7 million by 2027, Amália raises a question that resonates far beyond Lisbon: how does a country of ten million people deliver a sovereign LLM while France, despite greater resources and a world-class champion, still has no public national language model?
The Portuguese delivery in detail
According to the official Portuguese government statement, which presents Amália as the first open language model developed in European Portuguese, the project is the result of a consortium of Portuguese universities and research centers, involving more than sixty researchers. Coordination is led by NOVA University Lisbon, alongside Instituto Superior Técnico and the universities of Coimbra, Porto, and Minho, with support from the Foundation for Science and Technology (FCT). The model is made available as open source under the Apache 2.0 license, published on the Hugging Face platform under the amalia-llm organization.
The statement describes a model trained to understand text, documents, images, and speech, and adapted to the country’s language, legal context, and national reality. In practice, this multimodality is spread across several distinct models released by the team (a text model, a vision model, and a speech-recognition component), rather than being concentrated in a single system. The text core, a 9-billion-parameter model, was not trained from scratch: it builds on existing foundation models, including EuroLLM-9B, an open European multilingual model, as well as GlorIA, an earlier Portuguese model. The technical documentation for the released version states that it extends the pretraining of EuroLLM to better cover knowledge in European Portuguese and expands its context window to 32,000 tokens.
That clarification explains the budget. Adapting an existing foundation model costs an order of magnitude less than training one end to end, an operation that runs into tens or even hundreds of millions for state-of-the-art models. The initial €5.5 million, supplemented by an additional €1.5 million through 2027, therefore funds adaptation, data enrichment, the addition of multimodal capabilities, and the engineering work of several dozen people, on a largely shared European computing infrastructure. That is precisely what makes the project feasible on a modest public budget.
A method already proven elsewhere in Europe
The Portuguese case applies a formula that is now well established in Europe: start from an open existing base, adapt it to a national language and context, rely on European public supercomputers from the EuroHPC program, and release the result as open source. The Basque Country followed the same path with Latxa, an adaptation of Llama 2 to Euskara led by a university center. Spain went further with ALIA, a 40-billion-parameter model trained at the Barcelona Supercomputing Center, while Germany delivered Teuken-7B in late 2024, the result of the public OpenGPT-X consortium financed with roughly €14 million by the Federal Ministry for Economic Affairs.
These initiatives share an institutional architecture more than a benchmark score: a university consortium, targeted public funding, and shared European foundation and compute resources. At EU level, the OpenEuroLLM project, launched in February 2025 and bringing together around twenty organizations, aims to build a family of models covering all official languages; it has secured strategic access to several EuroHPC supercomputers. European linguistic sovereignty is thus being built through national building blocks backed by a common infrastructure, rather than through a single continental model.
The French contrast
It is against this backdrop that the French case stands out. France has a world-class champion, Mistral AI, whose Series C valued the company at around €11.7 billion in September 2025. But Mistral is a private company that releases open-weight models without being a public model funded and governed by the state. On the government side, the closest tool is Albert, developed by the Interministerial Directorate for Digital Affairs: a sovereign infrastructure that aggregates and serves third-party open models, including those from Meta and Mistral, on state servers, rather than a language model trained with public funds.
France does, however, have a notable precedent: BLOOM, a 176-billion-parameter multilingual large model trained in summer 2022 on the public Jean Zay supercomputer, as part of the international BigScience project coordinated by Hugging Face. Even so, BLOOM remained an international collective effort, with no ambition to serve as a dedicated national model for French. Since then, the French debate on sovereignty has been prolific — the Paris summit in February 2025, investment announcements, the third phase of the national strategy — yet it has not produced the kind of frugal, targeted equivalent that Amália represents: a public, open model adapted to the national language, delivered for the price of a research project.
The limits of the Portuguese feat
The achievement nonetheless has its limits. Amália remains an adaptation of an existing base, with the scope of a 9-billion-parameter LLM, far from leading American or Chinese systems. Some elements often repeated in the press also warrant caution: presenting Amália as an “alternative to American giants” reflects media framing more than government language, which speaks instead of sovereignty and transparency.
What matters is this: a small country has shown that a sovereign, open LLM adapted to its language can be delivered with a university consortium, earmarked European funding, and access to shared compute infrastructure. That demonstration, repeated from the Basque Country to Germany, leaves France facing a very concrete question. The Jean Zay supercomputer already trained BLOOM, EuroLLM is available under an open license, and Portugal has just put the operation at €7 million: all the ingredients for a public national model are on the table.