Another 1.8 Million Euros Carries Portugal's Home-Grown AI Model to December, and the Government Turns to Newspaper Archives for Training Data

Portugal's state-funded language model is getting another 1.8 million euros and six more months of development, and the government has said out loud what it will be shopping for next: the recent archives of the country's newspapers.

The money was signed off in Resolução do Conselho de Ministros n.º 178/2026 (Council of Ministers Resolution 178/2026), published on Tuesday in the Diário da República. It approves a second development phase for AMALIA, the Assistente Multimodal Automático de Linguagem com Inteligência Artificial (Automatic Multimodal Language Assistant with Artificial Intelligence), running to the end of the second half of 2026. The resolution was signed by Prime Minister Luís Montenegro on 4 September and takes effect retroactively from 1 July, so the phase it authorises is already two and a half months old. The 1.8 million euros comes from the Plano de Recuperação e Resiliência (Recovery and Resilience Plan), and according to ECO it lifts the model's total cost to about 7.3 million euros by the end of execution.

From chatbot to agent

The deliverables listed in the annex are more specific than the first phase's were. Alongside continued training of the core model, the government wants "agentification tools" for AMALIA: software that can be plugged into the internal processes of the public administration and into the delivery of public services, rather than sitting behind a chat window. The Navy was the first branch of the armed forces to put the model to work internally, in July. The stated ambition, as the government put it when it approved the spending on 9 September, is that AMALIA stops being a prototype and starts handling citizen enquiries about state services.

The annex also commits to further training on image and speech, and to expanding the model's coverage of Portuguese culture in partnership with five named institutions: the Cinemateca Portuguesa (Portuguese Film Archive), Património Cultural, I.P. (Cultural Heritage), Museus e Monumentos de Portugal, E.P.E. (Museums and Monuments of Portugal), the Direção-Geral do Livro, dos Arquivos e das Bibliotecas (Directorate-General for Books, Archives and Libraries) and the Biblioteca Nacional de Portugal (National Library of Portugal).

The part the journalists will notice

Separately from the resolution, the government announced that it intends to open a dialogue "shortly" with media companies to secure access to their more recent archives, in order to strengthen the model's training on Portuguese-language content. The Minister of the Presidency, António Leitão Amaro, framed it as Portugal setting an example in how a national model deals with news publishers, who he acknowledged "often raise concerns about the use of their archives".

There is history here. The first phase of AMALIA drew on Arquivo.pt, the national web archive, which harvests and preserves news published online in Portugal. The Sindicato dos Jornalistas (Journalists' Union) objected to that use at the time. Approaching publishers directly, and presumably paying or licensing, would be a different proposition from harvesting what a public archive has already captured.

Who actually runs it

The resolution also rewrites the governance. This phase is to be executed entirely by public bodies, with outside experts in an advisory role only. The Agência para a Reforma Tecnológica do Estado, I.P. (State Technological Reform Agency, or ARTE) takes responsibility for all of the funding, for coordination with university research centres, for running the inference infrastructure that serves the model, and for getting public bodies to actually use it.

The Fundação para a Ciência e Tecnologia (Foundation for Science and Technology) handles the data and the compute: access to Arquivo.pt and to Portugal's open-access scientific repositories, and the machines themselves, which means the Deucalion supercomputer, Portugal's share of MareNostrum-5 and its participation in the BSC AI Factory. Under Decreto-Lei n.º 132/2025, the foundation is due to be folded into a new body after merging with the national innovation agency, and the resolution anticipates that by naming "the entity that succeeds it".

The Specialist Monitoring Committee created in 2024 keeps its job, with its mandate extended all the way to 31 December 2029. It is chaired by Arlindo Oliveira, president of INESC, and includes Goretti Marreiros of the Portuguese Association for Artificial Intelligence, Daniela Braga of Defined.ai representing the Accelerat.ai consortium, and Paulo Dimas of the Center for Responsible AI. The resolution allows ministers to add members by order.

A third phase is already pencilled in. Point 6 states that AMALIA's development continues into 2027, on a timetable to be set by joint order of the finance, state reform and science ministers.

What this means for residents of Portugal

• You will probably meet AMALIA at a government counter before you meet it anywhere else. The agentification work is aimed at public service delivery, which in Portugal means the places foreign residents deal with most: tax, social security, registries.
• European Portuguese is the point. The stated objective is a model trained on European Portuguese from known and secure sources. If you have watched a general-purpose assistant answer you in Brazilian Portuguese, that is the gap this is meant to close.
• The money is small by AI standards. A total around 7.3 million euros is a fraction of what a commercial frontier model costs to train. AMALIA is a sovereignty and language project, not an attempt to compete on raw capability.
• Watch the publisher negotiations. How Portugal settles the question of paying newspapers for training data will matter well beyond this one model.

The state has now spent two years and roughly seven million euros building a Portuguese model it owns. The next six months decide whether it becomes something citizens actually use, or an expensive demonstration that the country can build one.