Portugal Launches Amalia Open Source Portuguese Language Model

Portugal's government unveiled Amalia on July 1, 2026 at Instituto Superior Tecnico's innovation center in Lisbon, with Prime Minister Luis Montenegro and ministers Fernando Alexandre and Goncalo Matias in attendance. The government describes Amalia, an acronym for Assistente Multimodal Automatico de Linguagem com Inteligencia Artificial, as the first open LLM built for European Portuguese. A consortium of more than sixty researchers, coordinated by NOVA University Lisbon alongside Instituto Superior Tecnico and the universities of Coimbra, Porto, and Minho, built it with Foundation for Science and Technology support, funded through Portugal's Recovery and Resilience Plan at EUR 5.5 million initially plus a further EUR 1.5 million earmarked through 2027. Alongside the launch, the government introduced IA.GOV.PT, the official AI portal for Portugal's public administration. "This is an important step in positioning Portugal in artificial intelligence, also showcasing the talent and knowledge of academia in service of society and the country," said Joao Magalhaes, a NOVA LINCS researcher and project coordinator, according to NOVA FCT's official announcement.

Rather than pretraining from scratch, the team extended the pretraining of EuroLLM-9B, an open European multilingual model, and drew on the earlier Portuguese GlorIA model, expanding the context window to 32,000 tokens. Per the team's own technical report, accepted at PROPOR 2026, the released roughly 9-billion-parameter text model, alongside a vision model and speech-recognition component, is published under the Apache 2.0 license on Hugging Face under the amalia-llm organization; the paper reports Amalia matches strong baselines on translated benchmarks while substantially improving on European-Portuguese-specific evaluations versus machine-translated alternatives.

Amalia extends a pattern already used elsewhere in Europe: the Basque Country's Latxa adapted Llama 2 for Euskara, Spain's ALIA trained a 40-billion-parameter model at the Barcelona Supercomputing Center, and Germany's Teuken-7B came out of the public OpenGPT-X consortium for roughly EUR 14 million. By contrast, France, home to Mistral AI (valued near EUR 11.7 billion in its September 2025 Series C), has no publicly funded national language model; its closest tool, Albert, aggregates third-party open models rather than training one. The EU-wide OpenEuroLLM project, launched February 2025 with about twenty member organizations and access to EuroHPC supercomputers, aims to formalize this adaptation-based approach across all official EU languages.

The reusable part of Portugal's approach is the adaptation pipeline, not the specific model: start from an open multilingual base like EuroLLM, extend pretraining and context length with targeted local-language data, and layer in modality-specific components, rather than funding a full pretraining run. That pipeline is what put a national LLM within reach of a university consortium and a research-grade budget instead of the tens-to-hundreds-of-millions typical of frontier training runs.

Whether Amalia sees real adoption in the public-service, healthcare, and education use cases the government has flagged; whether the OpenEuroLLM project's shared-compute model prompts other EU member states without a national LLM, starting with France, to follow the same adaptation playbook; and how Amalia's pt-PT-specific benchmark results, per the team's PROPOR 2026 paper, hold up against future frontier multilingual models.