AI & Tech · Lisbon

Portugal launches Amália, an open-source AI model for European Portuguese, with €7M investment

Portugal has officially released Amália, its first national AI model for European Portuguese, as open-source infrastructure. The €7 million project will power public services and local businesses without a chat interface.

AI text based on reports by SIC Notícias, SAPO, RTP and other publishers. Published . How it is made

For membersThe first 30 seconds for everyone. Synthetic voice (AI).

In brief

Lisbon · Portugal
How it came to this
  1. 1 JulOfficial launch of Amália with 9 billion parameters at Técnico Innovation Center.
  2. 31 DecPlanned general availability for all citizens and entities.
  3. 2027Next phase: expansion to 22 billion parameters and €1.5M additional funding.

Not a ChatGPT competitor

When the Portuguese government launched Amália on 1 July 2026, it made one point repeatedly: this is not a ChatGPT. Paulo Dimas, CEO of the Center for Responsible AI consortium, explained that there would be no chat interface for the public.

“This is not a ChatGPT. There won't be a chat interface for people to interact with, because that's not its function.”

Paulo Dimas, CEO of the Center for Responsible AI consortium

Instead, Amália is a foundation model that other software applications can call upon, a base layer for building services. It is a multimodal model that can process text, images and speech, and it has been optimized specifically for European Portuguese.

Open infrastructure for all

The model, its training data and source code are all released under an Apache 2.0 licence, making them free for commercial use by any company, university or government body. Users can download the model from the portal ia.gov.pt, which redirects to Hugging Face. CTO Manuel Dias described this openness as a deliberate strategic choice.

“This model has been validated with four entities – museums and culture, science, media, education. Now any company can take advantage of this model and develop solutions on top of it.”

Manuel Dias, CTO of the Portuguese State and president of ARTE

The licence ensures that the model's evolution is not controlled by a single large tech company, a contrast to proprietary systems.

Sovereignty and national data control

The launch was framed as a step toward digital sovereignty. Minister Gonçalo Matias said the government wants to reverse the trend of importing critical technology and instead build a national AI infrastructure that keeps citizens' data under Portuguese control.

“As we increase the digitalisation of our economy and society, we must guarantee control and sovereignty over our fellow citizens' data that we place in the digital world.”

Gonçalo Matias, minister of State and Digital Reform

The model is trained with European Portuguese datasets and built on EuroLLM-9B, a European foundation model, reflecting a broader EU push for sovereign AI.

Public services and early use cases

The government plans to integrate Amália into the gov.pt mobile app, where it could eventually alert citizens to expiring documents and offer to renew a citizen card. Other use cases demonstrated include an intelligent navigation system for the Portuguese Heritage 360 portal, an AI teaching assistant, a virtual guide for museums and monuments, and decision-support tools for the Portuguese Navy. The CTO stressed that Amália was not a lab project.

“Amália was not born to stay in the lab.”

Manuel Dias, CTO of the Portuguese State and president of ARTE

It is intended for deployment across public administration and the broader national ecosystem.

Funding and academic effort

The project was financed with an initial €5.5 million from Portugal's Recovery and Resilience Plan, with an additional €1.5 million announced at the launch, bringing the total to €7 million through 2027. The work took about 18 months and involved more than 60 researchers from five universities: NOVA Lisbon, Instituto Superior Técnico, and the universities of Porto, Minho and Coimbra. A test version was completed in September 2025 and presented at the PROPOR conference in Brazil. The next phase aims to scale the model from 9 billion parameters to 22 billion, increasing its capacity and performance.

Ask about this text

Sources and context

Places
Lisbon

Pollar is free and carries no ads in the text. Its members keep it going.

Support Pollar