Amália AI on Gov.pt: what Portugal’s artificial intelligence model represents

Artificial intelligence developed in Portugal is entering a new phase with the public release of the Amália model. According to information published by the organisations responsible for the project, the model can be accessed through the Gov.pt portal, which directs interested users to the Hugging Face platform, where the files required to use it can be downloaded.

This open release allows citizens, companies, public administration bodies, universities and research centres to explore a language model created with a very specific purpose: to understand and produce content in European Portuguese while retaining awareness of Portugal’s cultural, linguistic and institutional references.

Amália - an acronym for Automatic Multimodal Artificial Intelligence Language Agent - is presented as an open, public infrastructure. Rather than simply making another technological system available, the project seeks to strengthen national autonomy in a field dominated by major international companies and models trained primarily on English-language content.

Amália’s arrival on Gov.pt also raises important questions. What distinguishes this model from tools such as ChatGPT? What benefits can it offer Portuguese companies? What precautions should be taken when adopting it? And to what extent can an open model contribute to the economy, public services and the preservation of European Portuguese?

What is the Amália artificial intelligence model?

Amália is a large language model, commonly identified by the acronym LLM, which stands for Large Language Model. This type of technology learns statistical relationships between words, sentences, documents and other elements from large volumes of data. After training, the model can answer questions, summarise texts, produce content, classify information or support the performance of other language-related tasks.

In the first announced phase, Amália has approximately nine billion parameters. In simplified terms, parameters are the internal values adjusted by the model during training, enabling it to recognise patterns in language. The model is expected to evolve to 22 billion parameters, with the aim of increasing its capacity, accuracy and performance in more demanding tasks.

However, the number of parameters should not be interpreted in isolation as an automatic guarantee of quality. A model’s performance also depends on data quality, training diversity, alignment techniques, the architecture used, the tests conducted and its suitability for the context in which it will be applied.

For certain tasks, a smaller specialised model may produce more appropriate results than a larger model that is poorly adapted to the language or area of use. This is precisely where Amália seeks to establish its position: it does not aim to compete on size alone, but primarily through its specialisation in European Portuguese and its knowledge of the national context.

How can the model be downloaded and used?

According to the information available about the project, access to the model is provided through the artificial intelligence portal on Gov.pt. From that portal, users are directed to Hugging Face, a platform widely used by the artificial intelligence community to publish models, datasets, documentation and development tools.

The model was announced under an Apache 2.0 licence. This licence is commonly associated with open software projects and permits the use, modification and distribution of the published code or components, including in commercial contexts, provided that the conditions set out in the licence are met.

In practical terms, this means that a Portuguese company can download the model, evaluate it, adapt it to a specific sector and develop a commercial solution based on its capabilities. A public body can explore internal applications. A university can conduct scientific experiments. An independent developer can create prototypes or test new forms of interaction in European Portuguese.

The ability to download a model does not, however, mean that anyone can run it on a conventional computer. Requirements depend on the format provided, quantisation, the required memory and the type of task. Some models can be run locally on high-performance equipment, while others require servers equipped with graphics processing units or access to cloud infrastructure.

Why is an LLM focused on European Portuguese important?

A large proportion of the generative artificial intelligence models available on the market have been trained on much larger volumes of English-language content. Although these systems can communicate in Portuguese, they may produce less natural constructions, use Brazilian vocabulary, introduce unsuitable cultural references or struggle with legal, administrative and institutional terminology used in Portugal.

A model focused on European Portuguese can identify linguistic differences that may appear minor but have a significant impact on communication quality. Expressions, forms of address, spelling, administrative terminology and historical references form part of a country’s linguistic identity and influence the level of trust users place in a system.

For a company, this specialisation also has commercial value. A virtual assistant that responds in European Portuguese can provide a more consistent experience for customers in Portugal. A content production tool can reduce the need to correct Brazilian Portuguese constructions. An internal system may be better able to understand documents, regulations, contracts and communications produced in Portugal.

Linguistic adaptation can also support an SEO strategy aimed at searches carried out by Portuguese users. Although a language model cannot replace keyword analysis, editorial validation or a specialist’s expertise, it can support the creation of structures, frequently asked questions and content aligned with the vocabulary used by the national audience.

Preserving Portuguese culture and references

The Amália project is not limited to grammatical accuracy. One of its objectives is to preserve Portugal’s cultural representation. This includes proverbs, idiomatic expressions, public figures, monuments, historical events, works of art and other elements that may have limited representation in the datasets used by major international models.

The absence of these references is not merely an academic issue. A system that lacks knowledge of the national context may misinterpret a popular expression, confuse institutions, mix Portuguese and Brazilian realities or produce overly generic explanations of Portuguese heritage.

By incorporating data relevant to Portugal, Amália may become a tool for accessing culture and knowledge. Museums, libraries, archives, local authorities and educational institutions may explore new ways to organise, interpret and present information to different audiences.

This capability will become particularly important when the project advances towards a multimodal dimension. A multimodal model is not limited to text: it can analyse images, connect visual elements with descriptions and combine different forms of information in a single response.

Technological sovereignty and data protection

Another central objective of Amália is to strengthen sovereignty over citizens’ data. Many commercial artificial intelligence tools operate through external services, where requests are sent to infrastructure controlled by international providers. Depending on the service, configuration and contract, this use may raise questions about where information is located, processed, retained and protected.

The existence of an open model allows certain organisations to create private environments installed on their own infrastructure or hosted within national territory. This option may be relevant to public bodies, healthcare institutions, universities, industrial companies and other organisations that work with confidential information.

This does not mean that the model automatically guarantees data security. Technological sovereignty depends on the entire architecture: servers, networks, permissions, logs, backups, interfaces, integrations and internal practices. A locally installed model may remain vulnerable if the infrastructure is not properly protected.

Before using personal, strategic or confidential information, any organisation should carry out a risk assessment, define access rules and verify the relevant legal framework. There should also be clarity about the data used to adapt the model and whether that information could be reproduced in subsequent responses.

Supercomputers and European infrastructure

Training a large language model requires substantially more computing capacity than conventional software projects. According to the information released, Amália’s development relies on Portuguese and European supercomputers, including Deucalion and MareNostrum 5, as well as EuroHPC infrastructure.

These resources make it possible to process large volumes of text, run successive training cycles and test different configurations. The use of European infrastructure also places the project within the European Union’s broader strategy for technological autonomy.

Europe has sought to reduce its dependence on external platforms in critical areas such as cloud computing, semiconductors, data processing and artificial intelligence. The development of sovereign models in several countries shows that language, culture and legislation have become part of Europe’s technology strategy.

Portugal is therefore joining similar initiatives developed in other countries. Spain, Germany, Poland, the Netherlands, Denmark and Switzerland have projects focused on national or European languages and contexts. Each initiative has its own characteristics, but all share a concern with preserving scientific capacity and reducing strategic dependencies.

Applications of Amália in education

Education is one of the areas identified for the model’s application. Amália may support teachers and students through tools designed for the Portuguese educational context.

The announced uses include lesson planning, writing lesson summaries, creating study guides, supporting supervised study and generating practice tests. These tasks can reduce the time required to prepare materials and make it easier to adapt content to different learning levels.

A teacher may, for example, request suggestions for structuring a lesson on a particular subject, create questions with different levels of difficulty or adapt an explanation for students with different needs. A student may request a summary, exercises or an alternative explanation of a topic.

These possibilities require supervision. A language model does not understand content in the same way as a teacher and may provide incorrect information, invent references or produce incomplete explanations. Materials should be reviewed before they are used in a classroom or assessment process.

Artificial intelligence literacy is therefore becoming as important as the tool itself. Teachers and students need to learn how to formulate requests, verify sources, identify errors and distinguish technological assistance from validated knowledge.

Culture, heritage and museums

In the cultural sector, Amália may facilitate access to information about works of art, monuments and collections. The planned features include the annotation of visual elements, historical and artistic contextualisation, the identification of literary passages and the creation of semantic descriptions from images and textual attributes.

A museum may use these capabilities to prepare descriptions adapted to different audiences, create educational content or facilitate searches within a digital collection. An institution may connect a work with its historical period, artistic style and main influences.

The multimodal component may also support accessibility. The generation of image descriptions can assist visually impaired people, provided that the results are reviewed by specialists and integrated in accordance with accessibility best practices.

There is also a risk of error in this area. The identification of a work, figure or historical context should not depend exclusively on an automated response. Technology can support curators, researchers and communication teams, but it does not replace the scientific work required to validate information.

Media, journalism and information analysis

In the media sector, the project includes tasks such as abstractive source summarisation, the identification of dominant narratives, the detection of persuasion techniques, the generation of news reports from structured data and the creation of photograph captions.

Abstractive summarisation does not simply copy sentences. The model interprets the content and produces a summary using different wording. This capability can help journalists and analysts review lengthy documents, minutes, reports or collections of articles.

Another proposed application involves analysing potentially manipulative content. A system can search for linguistic patterns associated with particular persuasion techniques and present a justification. However, classifying a text as manipulative requires caution, context and transparent criteria.

The automated production of news reports from sports statistics or municipal meeting minutes can also accelerate repetitive tasks. Even so, editorial supervision, clear disclosure of artificial intelligence use and fact-checking are required. An accurate news report involves more than converting figures into sentences: it requires context, relevance, balance and responsibility.

Public trust will depend on how these technologies are used. Media organisations should define internal policies covering transparency, authorship, review, source protection and error correction.

Scientific research and access to knowledge

In science, Amália may support searches across theses, academic papers, historical documents and other Portuguese-language sources. The model is expected to summarise scientific literature, answer questions in different fields and provide explanations of the process supporting an answer.

For researchers, the ability to search large Portuguese-language collections may facilitate the discovery of studies that are less visible in international databases. It may also help compare documents, identify recurring themes or prepare an initial organisation of the literature.

However, a language model may invent titles, authors, quotations or findings. Any suggested reference must be checked against the original source. In a scientific context, a plausible answer is not the same as a correct answer.

The openness of the model may have an additional effect: it allows Portuguese universities to study how it works, propose improvements and develop specialised versions. This collaboration may help train researchers and engineers with practical experience of large-scale models.

Is Amália equivalent to ChatGPT?

Amália should not be compared directly with ChatGPT. ChatGPT is an application that uses models from the GPT family and includes an interface, security systems, tools, memory, integrations and other components. Amália is primarily presented as a language model that can provide the foundation for different applications.

A company may use Amália to create its own assistant, but it will still need to develop or integrate the remaining components: interface, authentication, document search, storage, permission controls, monitoring, security mechanisms and connections to internal systems.

There are also differences regarding openness. The Portuguese project plans to make the model publicly available together with a model card containing technical information about its characteristics, limitations and training data. As a rule, closed commercial models do not allow the same level of inspection or adaptation.

On the other hand, a large-scale commercial application benefits from years of development, vast computing resources and teams dedicated to security, user experience and operations. Releasing an open model is an important step, but turning that model into a robust product requires additional investment.

Opportunities for Portuguese companies

The open licence may encourage Portuguese companies to create solutions adapted to specific sectors. Tourism, industry, retail, banking, insurance, education, healthcare, energy, legal services and public services are some of the areas in which a European Portuguese model could be explored.

A company can adapt the system to answer questions based on its documentation, classify requests, summarise reports or support internal teams. It can also create search tools that consult manuals, catalogues, procedures and knowledge bases.

In e-commerce and Shopify integrations, a language model can support the creation of product descriptions, catalogue organisation, responses to frequently asked questions and natural-language search. All of these uses should include human validation and rules designed to prevent incorrect answers about prices, availability or commercial terms.

The main competitive advantage will not lie solely in access to the model. As the technology is open, other organisations will also be able to use it. Differentiation will result from data quality, sector knowledge, integration with real processes and the experience provided to users.

Downloading a model does not mean having a complete solution

The release of Amália may create the impression that downloading the files is enough to obtain a complete assistant. In reality, an enterprise artificial intelligence project includes several stages.

1. Define the problem: identify a specific task, the users involved and the expected outcome.
2. Assess the data: verify the quality, permissions, confidentiality and format of the available information.
3. Choose the architecture: decide between local use, cloud infrastructure, API access or a hybrid solution.
4. Integrate systems: connect the model to the website, store, CRM, internal search or other platforms.
5. Create security mechanisms: limit access, filter content and prevent the exposure of sensitive data.
6. Test and monitor: measure accuracy, costs, speed, satisfaction and types of error.

In many cases, it will not be necessary to retrain the entire model. Techniques such as retrieval-augmented generation, known as RAG, make it possible to provide the system with relevant documents before it produces a response. This approach can reduce costs and make information easier to update.

Quantised versions can also be used to reduce memory requirements, although they may cause some loss of quality. The choice should be based on testing rather than technological preference alone.

Risks, incorrect answers and human oversight

Like any large language model, Amália may produce incorrect responses. These systems generate text based on patterns and probabilities. They do not possess an internal guarantee of truth and may present false statements in a convincing tone.

For this reason, the organisations responsible for the project have highlighted the need to inform users that responses were generated by artificial intelligence. This transparency is particularly important in education, healthcare, justice, public services, finance and other areas with a significant impact on people.

Organisations should establish the situations in which a response requires human validation. They should also allow users to challenge a result, report an error or request assistance from a person.

Testing should include offensive language, ambiguous questions, manipulation attempts, personal data and content outside the intended area. A model’s average quality does not necessarily reveal how it will behave in the most sensitive cases.

Bias and alignment techniques

Models learn from data produced by people and may reproduce inequalities, stereotypes or distortions found in that content. The Amália project includes alignment techniques intended to reduce bias and limit the generation of harmful content.

Alignment may include instruction data, human evaluation, filters, rules and guardrails. None of these mechanisms eliminates risk entirely. In addition, protection that is too restrictive may prevent legitimate responses, while insufficient protection may permit inappropriate content.

The work must be continuous. As the model is used, new examples, failures and forms of interaction will emerge that were not previously anticipated. A responsible policy requires monitoring, documentation of changes and channels through which the community can provide feedback.

Copyright and training data

According to the information released, pre-training used open data, primarily from the Web, social networks and sources with a high level of linguistic quality. Content extracted and filtered from Arquivo.pt was also used, with data explicitly stating that it should not be used for technical or scientific purposes excluded from training.

Intellectual property is one of the most widely debated issues in generative artificial intelligence. The fact that content is accessible on the Internet does not necessarily mean that it can be used without restrictions. Licences, copyright, database rights and research exceptions may vary according to the source and intended purpose.

A public project should seek a high degree of transparency regarding data categories, exclusion criteria and the mechanisms available to rights holders. The publication of technical documentation can help researchers and users understand the model’s limitations.

Companies that adapt Amália also assume responsibilities. The use of internal documents, customer content or protected materials must comply with contracts, licences, privacy rules and intellectual property rights.

Framework under the European Artificial Intelligence Act

The project presents the model as general-purpose artificial intelligence without systemic risk. Under the European Artificial Intelligence Act, the level of risk depends not only on the base model but also on the specific application.

A system used to create draft text does not necessarily present the same level of risk as a tool that influences decisions about employment, credit, education or access to essential services. The organisation developing or providing the application must assess its purpose, data, users and impact.

Companies should not assume that using a public model removes their legal obligations. Depending on the case, duties may exist regarding transparency, documentation, oversight, risk management, data protection and user information.

Involving legal, technical and operational teams from the beginning reduces the likelihood of creating a functional solution that cannot be used responsibly.

From a text-based version to a multimodal model

Amália’s planned development includes modalities beyond text. A multimodal model can receive images, text and other formats, establishing relationships between different types of content.

This capability creates opportunities in museums, education, accessibility, communication and research. The system may describe an image, identify visual elements, relate them to textual information or generate content from a combination of sources.

At the same time, risks increase. An incorrect visual interpretation may lead to false identifications. Images may contain personal data, faces, documents or sensitive information. Assessments must cover every modality used.

Multimodality should not be treated merely as an additional feature. It requires new training data, metrics, filters, tests and explanation mechanisms.

Economic and scientific impact for Portugal

Amália may contribute to the creation of a national artificial intelligence ecosystem. Its open release allows universities, companies and public administration bodies to work from a common foundation, share experiences and develop applications adapted to Portuguese reality.

This ecosystem may create demand for machine learning engineers, data scientists, language specialists, cybersecurity professionals, legal experts, experience designers and business consultants. It may also stimulate companies that provide model hosting, integration, evaluation and adaptation services.

The existence of national technology also helps retain knowledge. When all solutions depend on external platforms, a significant proportion of the expertise, data and economic value remains concentrated outside the country.

However, sovereignty cannot be achieved simply by releasing a model. It requires continued funding, infrastructure, stable teams, technical updates and genuine adoption by organisations.

What should interested organisations do?

The first step should not be to download the model, but to select a measurable use case. An organisation can begin with a low-risk internal task, such as searching non-confidential documentation or creating drafts that remain subject to review.

It should then build a test set containing real questions and expected answers. This assessment makes it possible to compare Amália with other options and understand where its specialisation in European Portuguese provides an effective advantage.

It is also important to calculate costs. Even an open model requires infrastructure, storage, development, security and maintenance. In some cases, an external service may be more economical. In others, a private installation may be justified by confidentiality requirements or the volume of use.

The choice does not need to be exclusive. An architecture can combine different models according to the task, risk level and language. Amália may play a central role in processes that benefit from Portuguese context, while other systems may be used in different scenarios.

An open infrastructure for creating Portuguese applications

The public release of Amália represents an important step for artificial intelligence in Portugal. For the first time, companies, researchers, public bodies and citizens have access to a large-scale national model focused on European Portuguese and designed to evolve through collaboration.

The project’s value will be measured by the quality of the applications that emerge, the rigour of its documentation, its ability to correct failures and its real impact on services. An open model can democratise experimentation, but trust will depend on how it is implemented.

Education, culture, media and science are the first identified areas, but the possibilities extend across numerous economic activities. Linguistic specialisation, operation within private environments and freedom to adapt the model may make Amália a relevant foundation for solutions created in Portugal.

The challenge now is to transform technological capacity into useful, secure and accessible products. This requires cooperation between research, business, public administration and civil society, together with continuous assessment of results and risks.

BYDAS supports companies with the responsible integration of artificial intelligence into digital projects, from content and automation to digital marketing platforms. A well-defined strategy can transform models such as Amália into useful, measurable solutions aligned with each organisation’s business objectives.