AMÁLIA is no longer a prototype. Council of Ministers Resolution no. 178/2026, published in the Diário da República (Portugal’s official gazette) on 15 September, approves the second phase of the public artificial intelligence model — a further €1.8 million from the PRR, Portugal’s EU Recovery and Resilience Plan, backdated to 1 July — and gives it a concrete purpose: supporting citizen services across the State.
Until now, the question was whether the public administration would adopt AI. That question has been answered — on paper. The one that remains is different: with which model, trained on which data, and with what degree of dependence on external suppliers. The resolution answers with a bet: a model of its own, open under the Apache 2.0 licence, trained in European Portuguese and, in the resolution’s own terms, on data of known origin.
Two milestones to note. The second phase runs to the end of 2026, with multimodal capabilities (text, voice and image) and agentification tooling — in practice, AI that carries out tasks within the administration’s processes rather than merely answering questions. Development then continues into 2027, with guidance to be set by ministerial order; cumulative investment comes to around €7.3 million.
If you deal with the State
AMÁLIA came out of a consortium of public universities, coordinated by the FCT (the national science funding agency) and ARTE (the State Technological Reform Agency), and the first version shipped in July — with its performance debated in the press, hallucinations included. It is a model under construction, and the resolution itself acknowledges as much by funding further training. The direction, though, has been chosen: the resolution wants the answer given to a citizen or a company at a public service desk to be prepared by a model the administration operates, with the data under its control. With external services that control is weaker — and it is precisely that dependence the bet refuses. And AI-assisted public service is personal data processing: who answers for that processing, and on what legal basis, is a discussion this phase’s regulation will have to settle.
For now, nothing demands action. The signal to watch for is assisted service appearing in the public services you use — and the data rules that come with it.
If you build technology
An open, national model trained in European Portuguese is also raw material — and it is already being used that way: Caravela, by Podsqueeze, embeds AMÁLIA as a component of its AI infrastructure. Here, “open” has a precise definition: the Apache 2.0 licence, which allows the model to be downloaded, modified and built into commercial products with no licensing costs — publicly confirmed by the State’s chief technology official. In numbers: around 9 billion parameters, built on EuroLLM-9B, with a 32,000-token context window.
For anyone processing language in a Portuguese context — documents, customer service, search — choosing a model is no longer about picking the top of the benchmark table. It is about deciding who you become dependent on, where the data ends up, and what growth will cost.
The next step fits in a sentence: download AMÁLIA from the ia.gov.pt portal and test it against your own use case before putting it on the roadmap. What follows has the most variables per case — performance, compliance and dependence rarely have the same answer for any two companies. It is at that crossing of the technical and the legal that Anjos & Brum works, and where a timely conversation saves an untimely migration.
Informational content as at 21/09/2026; each case deserves its own analysis.