Portugal Delivers Open LLM Amália for €7M While France Still Debates AI Sovereignty

On July 1, 2026, the Portuguese government presented Amália, which it calls the first open large language model (LLM) developed in European Portuguese. The project, coordinated by NOVA University of Lisbon with other universities and research centers, mobilized over sixty researchers and was funded by the Recovery and Resilience Plan (PRR) with a public investment rising to €7 million by 2027. The model is available under Apache 2.0 license on Hugging Face. Amália is not a single system but a set of models: a 9-billion-parameter text model, a vision model, and a speech recognition component. The text model builds on existing open models, notably EuroLLM-9B and GlorIA, extending EuroLLM's pretraining to improve European Portuguese knowledge and increasing context window to 32,000 tokens. This approach cost significantly less than training from scratch. Similar strategies have been used elsewhere in Europe, such as the Basque Latxa, Spain's ALIA, Germany's Teuken-7B, and the EU's OpenEuroLLM project. In contrast, France has Mistral AI, a private company, and Albert, a state infrastructure that serves third-party open models, but no publicly trained national language model, despite previous efforts like BLOOM. Amália has limitations: it is an adaptation of an existing base and only 9 billion parameters, far from top-tier US or Chinese systems. Nevertheless, it demonstrates that a sovereign, open LLM adapted to a language can be achieved with a university consortium, European funds, and shared computing infrastructure.