Portugal launches Amália, its first sovereign open-source AI model — for 7 million euros

Portugal launches Amália, its first sovereign open-source AI model — for 7 million euros

🔎 While France debates, Portugal delivers

Europe has a chronic problem with artificial intelligence: it talks a lot, it acts little. While France chains together parliamentary missions, strategic reports, and debates on "digital sovereignty," Portugal simply built a model. And released it.

Amália, named after the fado legend Amália Rodrigues, is the first open-source language model designed specifically for European Portuguese. Not an adaptation. Not a superficial fine-tuning of an English-language model. An LLM thought up from scratch to capture the linguistic and cultural nuances of Portuguese as it is spoken in Portugal.

The signal is as much political as it is technical. At a time when the United States dominates the LLM wave with models like Gemini 3.1 Pro, GPT-5.5 or Claude Opus 4.7, and China lines up competitors like DeepSeek V4 Pro, Europe is desperately searching for its place. Portugal just gave it one.

The most striking part? The price. Seven million euros. That is ridiculously low compared to American budgets. And that is exactly the point.

The Essentials

• Amália is Portugal's first sovereign open-source LLM, developed for European Portuguese by a consortium of universities with €5.5M in European funding and a total budget of €7M.
• The model, its training datasets, and its source code are published under an open-source license — zero restrictions, zero barriers.
• Priority targeted applications: public services, academic research, and the Portuguese administration.
• This launch is part of the European push for AI sovereignty, alongside initiatives like the NVIDIA Nemotron 3 Ultra 550B or European open-weights models.
• The Portuguese strategy contrasts with the French approach: less debate, more delivery.

Tools and models mentioned

Who is behind Amália — a consortium, not a startup

Amália is not the product of a Portuguese unicorn that raised 100 million euros. It is the result of a consortium of Portuguese universities and research institutions, directly supported by the government.

According to Reuters, the model was developed as part of a coordinated effort involving several research centers, with INESC-ID as the acknowledged lead. The stated goal: "to give a voice to Portuguese identity through artificial intelligence."

The funding is twofold. On the one hand, 5.5 million euros in European funds, according to Technology.org. On the other, a national complement bringing the total budget to around 7 million euros, as reported by ActuIA.

This hybrid funding model (Europe + Member State) could become a template. It shows that you don't need a Mistral AI to produce a sovereign model. Well-organized and properly funded researchers are enough.

This is a fundamental difference from the French ecosystem, dominated by the public-private duo Mistral/LightOn. Portugal chose the purely academic path. The result is less spectacular in terms of benchmarks, but potentially more sustainable institutionally.

Why European Portuguese needed its own model

Most current models, even those that "speak" Portuguese, are trained primarily on Brazilian Portuguese. This is a major problem for a sovereign state.

European Portuguese and Brazilian Portuguese share a grammatical base, but diverge deeply in vocabulary, syntax, idiomatic expressions, and cultural references. It's a bit like confusing French from France with Quebec French — except that the gap is even more pronounced.

As Duarte Carmo explains, AMALIA is designed as "a fully open-source LLM for European Portuguese" whose goal is to process this language "natively". Not as an add-on. Not as a secondary dialect.

The nuances captured by Amália include the grammatical specificities of continental Portuguese, Portuguese administrative and legal vocabulary, and cultural references specific to the country. According to INESC-ID, the model aims to "understand, process and generate content in European Portuguese, capturing linguistic and cultural nuances".

This issue is not unique to Portugal. It concerns all European languages that are not English. This is exactly the reason why rankings like the meilleurs LLM en français exist — because the quality of a model depends directly on its native processing of the target language.

What is truly open-source — and what it implies

The term "open-source" has become a buzzword in AI. Meta calls Llama "open-source" even though its license restricts commercial use beyond 700 million monthly users. Google does the same with Gemma. The line between open-source, open-weights, and source-available has become blurred — a topic we covered in detail with the case of Meta Muse Spark.

Amália, on the other hand, plays the total transparency card. According to Zonebourse and Euronext, three elements are published: the model itself, the training datasets, and the source code. Under a genuine open-source license.

This means that any researcher, company, or government agency can: download the model, inspect exactly what it was trained on, modify it, retrain it, deploy it locally. For Portuguese public administrations, this is a guarantee of absolute sovereignty. No data passes through American servers.

For those who want to go further with local model running, our local LLM installation guide details the steps with Ollama and LM Studio — tools perfectly suited for deploying this type of sovereign model on your own machines.

7 million euros: cheap or effective?

Seven million euros for a sovereign LLM is a drop in the ocean compared to American budgets. GPT-5.5 likely cost hundreds of millions in compute alone. But comparing Amália to GPT-5.5 is like comparing a utility vehicle to a Ferrari. They don't address the same need.

Amália's budget covers: the collection and cleaning of datasets in European Portuguese, model training, fundamental research, the salaries of academic researchers, and computing infrastructure. According to Cybernews, this budget is considered optimized for a model with a national purpose.

The effectiveness lies in the target. Amália doesn't need to beat Gemini 3.1 Pro on English-language benchmarks for advanced mathematics. It must excel at European Portuguese in administrative, legal, and cultural contexts. It's a specialized model, not a generalist model.

This approach echoes the philosophy behind the best LLMs for research: specialization often beats generalization when the use case is well-defined. Perplexity doesn't try to compete with GPT-5.5 on poetry — it optimizes for factual research. Amália optimizes for European Portuguese.

The value for money is therefore potentially excellent. Seven million for a tool that public administration can deploy without recurring licensing fees, without foreign dependency, and without data leaks. In the context of digital sovereignty, it's a modest investment with structural returns.

The concrete applications targeted by Portugal

Amália is not an academic exercise with no real-world outlet. The Portuguese government has specific use cases in mind.

According to Portugal Resident, initial applications are planned for public services. Specifically, this covers: modernizing citizen-State interfaces, translating and drafting administrative documents in correct European Portuguese, supporting civil servants in processing files, and analyzing national legal corpora.

OpenSourceForU points out that the model also targets researchers and businesses, not just the State. The idea is to create an ecosystem around Amália: startups building specialized applications on top of the base model, labs improving it, and government agencies customizing it.

This is the "sovereign platform" model. Portugal is not looking to compete with OpenAI or Anthropic in the consumer market. It is building the base infrastructure on which its local ecosystem can develop.

For Portuguese developers who want to build AI agents on top of Amália, the architecture is compatible with the approaches described in our article on the best LLMs for AI agents. The model can serve as a reasoning base for local agentic chains.

European AI sovereignty: where do we really stand?

Amália is part of a broader movement. Europe has been pushing for AI sovereignty since the launch of the AI Act regulation, but concrete results remain scattered.

On the large open-source model front, Europe has a few cards to play. The NVIDIA Nemotron 3 Ultra 550B, although a US model, represents the level of performance achievable in open-source. In France, Mistral continues to release competitive models, although the tension between open-weights and true open-source persists.

Other European academic initiatives are emerging. Models like GLM-5.1 (Z.AI) achieve scores of 83 in the overall ranking, proving that non-American research can produce credible results. In terms of architectural innovation, Google's DiffusionGemma has shown that the diffusion approach can be 4x faster than autoregressive — a path that European projects could explore.

But the fact remains: Europe is fragmented. Each country is pulling in its own direction. Portugal with Amália, France with Mistral, Germany with its own initiatives. There is a lack of large-scale European strategic coordination.

The parallel with Sentient Foundation, which raised 42 million dollars for open-source AI, is illuminating. US private funding for open-source AI far exceeds European public funding. But Amália proves that public money, when well-targeted, can produce concrete results.

The France-Portugal parallel: debates versus delivery

The contrast is striking. France has produced more reports on digital sovereignty than sovereign models deployed within its public administration. Portugal, with infinitely more modest means, has delivered a functional, documented, and open model.

This does not mean the French strategy is empty. Mistral AI is an undeniable success in terms of global tech competition. But Mistral is a private company whose most powerful models are now closed or under restrictive licenses. The sovereignty of a publicly traded startup is not the sovereignty of a State.

Portugal made a different choice: entrusting sovereignty to public research, funding it with public and European money, and publishing the result without restrictions. It is a model of digital sovereignty through academia, not the market.

For a country of 10 million inhabitants, it is a bold but rational gamble. Portugal does not have the critical mass to create a Google or an OpenAI. But it does have the critical mass to create an excellent model in European Portuguese — a niche that no one else is going to serve properly.

The lesson for France and for Europe is clear: sovereignty is not decreed, it is built. And sometimes, seven million euros well spent are worth more than a hundred million in conditional subsidies.

How Amália positions itself against current leaders

To be honest: Amália is not going to dethrone the leaders of the LLM leaderboard. Gemini 3.1 Pro dominates with a score of 92, followed by GPT-5.5 at 91 and Claude Opus 4.7 at 90. In agentic, GPT-5.5 reaches 98.2 — a level of performance that belongs to a completely different investment category.

But these comparisons are beside the point. Amália is not competing with GPT-5.5. It is competing with the absence of a model in European Portuguese. And on this specific ground, it has no competitor.

The real comparison would be: what happens when you ask GPT-5.5 or Claude Opus 4.7 to draft an administrative document in European Portuguese? The answer is well known — these models mix variants, produce Brazilian phrasing, and lack cultural nuances. Amália is built precisely to avoid these pitfalls.

For French-speaking users, the situation is analogous. Even the best free LLMs like ChatGPT Free or Gemini still make glaring errors in French — anglicisms, calqued phrasing, and off-base cultural references. This is the exact problem that Amália solves for Portuguese.

The limits of the model — what it doesn't do

A sovereign model costing 7 million euros necessarily has limits that need to be made explicit.

Firstly, the size. The sources do not detail the exact number of parameters, but a €7M budget imposes severe compute constraints. It is likely far from the 550 billion parameters of Nemotron 3 Ultra. Amália is more likely in the 7B-70B range, optimized for linguistic quality rather than raw reasoning capacity.

Secondly, relative monolingualism. A model optimized for European Portuguese will be less performant in English, French, or other languages. For multilingual tasks, generalist leaders remain indispensable.

Thirdly, the ecosystem. An open-source model without an active community dies quickly. Amália's success will depend on adoption by Portuguese developers, external contributions, and the applications that will be built on top of it. The government can launch a model, but it cannot force an ecosystem.

Finally, maintenance. An LLM is not static software. It ages, its knowledge depreciates, and the linguistic landscape evolves. Who will fund the updates in two years? Five years? This is the question that every sovereign project must face.

For technical teams that would like to experiment with Amália in addition to other models, our page on the meilleurs LLM locaux offers an overview of personal machine deployment options — an approach perfectly suited to an open-source model like Amália.

❌ Common mistakes

Mistake 1 : Confusing open-weights and open-source

Many media outlets present Meta or Google models as "open-source" even though their licenses include significant restrictions. Amália, on the other hand, publishes the model + datasets + code under a true open-source license. The distinction is not anecdotal — it determines what you can legally do with the model.

Mistake 2 : Judging Amália on English-language benchmarks

Evaluating this model on MMLU or HumanEval in English makes no sense. It's like rating a Japanese chef on his ability to make a beef bourguignon. The relevant criterion is the quality in European Portuguese, in Portuguese administrative and cultural contexts.

Mistake 3 : Believing that €7M is enough for everything

The budget is impressive for what it has produced, but it does not cover nationwide deployment infrastructure, long-term maintenance, or the development of final applications. Amália is a foundation, not a finished building.

Mistake 4 : Ignoring the dependence on American compute

Even a European sovereign model likely depends on NVIDIA GPUs for training. True sovereignty also requires hardware independence — an issue that Europe has not yet resolved.

❓ Frequently Asked Questions

Is Amália truly open-source or open-weights?

Truly open-source. Unlike Llama or Gemma, which only publish weights under a restrictive license, Amália publishes the model, training datasets, and source code under an open-source license, according to Reuters and Zonebourse (July 2026).

Can Amália be used outside of Portugal?

Yes. The open-source license is not geographically restricted. However, the model is optimized for European Portuguese and Portuguese cultural contexts. Its usefulness for other languages or regions will be limited.

Can Amália replace GPT-5.5 or Claude for general use?

No. It is a model specialized in European Portuguese. For general multilingual use, leaders like Gemini 3.1 Pro, GPT-5.5, or Claude Opus 4.7 remain vastly superior in terms of reasoning and versatility.

How to deploy Amália locally?

Being open-source, the model can be deployed with tools like Ollama or LM Studio, as detailed in our local LLM installation guide. The exact hardware specifications depend on the model size, which was not yet detailed at the time of launch.

Why Portugal and not another European country?

Portugal benefits from a favorable combination: a language with a strong identity (distinct from Brazilian), a solid research ecosystem (INESC-ID, universities of Lisbon and Porto), and a clear political will. Other European countries have the means but not the same perceived urgency.

✅ Conclusion

Portugal spent 7 million euros and achieved what countries ten times richer have not yet produced: a sovereign language model, truly open-source, optimized for its national language, and deployable without foreign dependency. Amália is not going to change the global AI game. But it changes the game of what "European digital sovereignty" means in practice — and it shows that delivery is better than debate. To follow the evolution of open-source models like Amália in the European landscape, check out our comparison of the best open-weights LLMs.