Amalia LLM Review 2026: Portugal’s New 9B Model

Portugal has long been known for its rich cultural heritage, from the haunting melodies of fado to the historic architecture of Lisbon and Porto. But in 2026, the country has made a significant entry into the global AI landscape with the launch of Amalia, a fully open-source large language model purpose-built for European Portuguese (pt-PT).

Named after Amália Rodrigues, the iconic fado singer whose voice defined a generation, Amalia is more than just a linguistic tool—it is a cultural statement. The name itself is an acronym: Assistente Multimodal Automático de Linguagem com Inteligência Artificial (Automatic Multimodal Language Assistant with Artificial Intelligence).

In this review, we take a deep look at what makes Amalia special, how it stacks up against its predecessor EuroLLM-9B, and whether it deserves a spot in your AI toolkit.

What Is Amalia LLM?

Amalia is a 9-billion parameter language model developed by a consortium of Portuguese researchers and institutions. It was built on top of EuroLLM-9B, an existing open-source European model, but significantly enhanced with additional high-quality training data specific to European Portuguese.

According to the team’s technical report (arXiv:2603.26511), Amalia was designed to address a critical gap: while many large language models handle Portuguese, they often conflate pt-PT (European Portuguese) with pt-BR (Brazilian Portuguese). This distinction matters. European Portuguese has distinct phonetics, vocabulary, and cultural nuances that machine-translated benchmarks frequently miss.

Amalia treats the pt-PT / pt-BR distinction as a first-class evaluation dimension, ensuring that the model doesn’t just “speak Portuguese” but speaks it authentically for its target audience.

How Amalia Was Built

Amalia’s development followed a two-stage approach:

1. Mid-training: The model was exposed to a curated corpus of high-quality European Portuguese text, including literature, news, academic papers, and conversational data.
2. Post-training: The team applied Direct Preference Optimization (DPO) to refine the model’s output quality, aligning it more closely with human preferences.

The result is a model that not only matches strong baselines on translated benchmarks but substantially outperforms them on European-Portuguese-specific evaluations. This is a meaningful achievement, as many models appear competitive on paper but falter when evaluated on native data.

Key Features and Capabilities

European Portuguese Expertise

Amalia’s primary selling point is its deep understanding of pt-PT. This includes:

• Vocabulary and idioms: The model recognizes expressions unique to European Portuguese that might be lost in translation.
• Cultural references: From fado to football, Amalia understands the cultural context of its target audience.
• Grammar and syntax: European Portuguese has distinct grammatical features (such as the use of “tu” vs. “você”) that Amalia handles natively.

Open Source and Accessible

Amalia is fully open source, meaning developers, researchers, and businesses can use, modify, and distribute it freely. The model is available on Hugging Face under the repository amalia-llm/AMALIA-9B-0626-DPO, making it easy to access and integrate into existing workflows.

Multimodal Potential

While the initial release focuses on text, the “Multimodal” in Amalia’s name hints at future capabilities. The consortium has indicated that multimodal extensions—particularly for image and audio processing—are on the roadmap, which could make Amalia a versatile tool for a range of applications.

Fairness and Misinformation Awareness

Like any LLM, Amalia can produce text that is false, misleading, or harmful. However, the team has explicitly addressed misinformation and misuse as design considerations, building in safeguards to reduce the risk of generating inaccurate or biased outputs.

Performance Benchmarks

Amalia’s performance is best understood through two lenses:

1. Translated benchmarks: On standard benchmarks that have been translated into Portuguese, Amalia matches or slightly exceeds strong baselines.
2. Native pt-PT evaluations: On evaluations specifically designed for European Portuguese, Amalia shows substantial improvement over machine-translated alternatives, confirming that its targeted training approach pays off.

This dual performance profile is significant. Many models appear strong on translated data but struggle with the subtleties of native language use. Amalia’s strength on native evaluations suggests it will perform well in real-world Portuguese-language applications.

Amalia vs. EuroLLM-9B

Since Amalia is built on EuroLLM-9B, it’s worth comparing the two:

Feature	EuroLLM-9B	Amalia LLM
Base model	—	EuroLLM-9B
Training data	General European	Curated pt-PT focused
pt-PT vs pt-BR	Handled but not prioritized	First-class evaluation dimension
Post-training	Standard	DPO-enhanced
Open source	Yes	Yes
Best for	General European Portuguese	Deep pt-PT expertise

The key takeaway: if you need a model that works well for Portuguese in general, EuroLLM-9B is solid. But if you need authentic European Portuguese performance, Amalia is the superior choice.

Pricing and Availability

As a fully open-source model, Amalia is free to use for most purposes. The consortium has made the model accessible through Hugging Face, and there are no licensing fees for commercial use.

For organizations that prefer managed hosting or enterprise support, several cloud providers offer Amalia through their AI marketplaces. Pricing for managed services varies by provider but typically follows standard per-token or per-call models, making Amalia a cost-effective option for businesses of all sizes.

Who Should Use Amalia?

Amalia is particularly well-suited for:

• Portuguese-language content creators who need authentic pt-PT output
• Businesses operating in Portugal that require culturally aware AI for customer service, marketing, and internal communications
• Researchers and developers working with European Portuguese data
• Educational institutions looking for a model that respects linguistic nuances
• Any organization that values open-source AI and wants to support European linguistic diversity

Pros and Cons

Pros

• Authentic pt-PT performance: Superior handling of European Portuguese nuances
• Fully open source: Free to use, modify, and distribute
• Built on proven foundation: Leverages EuroLLM-9B’s strengths
• Culturally aware: Understands references and expressions specific to Portugal
• Transparent development: Technical report and research papers available
• Growing ecosystem: Active community and roadmap for multimodal extensions

Cons

• Newer model: Less long-term track record compared to established LLMs
• pt-PT focus: May be less optimal for Brazilian Portuguese applications
• 9B parameter size: Smaller than some competitors, which may limit performance on complex tasks
• Limited multimodal support: Initial release is text-focused

Use Cases and Applications

Amalia’s capabilities make it suitable for a wide range of applications:

• Customer service chatbots that need to communicate naturally with Portuguese-speaking customers
• Content generation for Portuguese-language media, including news, blogs, and social media
• Educational tools that require accurate Portuguese grammar and cultural references
• Legal and administrative documents where precision in pt-PT is critical
• Translation and localization workflows that benefit from native-level Portuguese understanding

Looking Ahead

The Amalia consortium has announced plans for multimodal extensions, which could add image and audio processing capabilities to the model. This would position Amalia as a more versatile tool for applications that require understanding beyond text.

Additionally, the team is working on domain-specific fine-tunes for sectors such as healthcare, law, and education, which could further enhance Amalia’s utility for specialized applications.

Frequently Asked Questions

What is the difference between Amalia and EuroLLM-9B?

Amalia is built on EuroLLM-9B but features enhanced training data focused on European Portuguese, DPO post-training, and a first-class treatment of the pt-PT / pt-BR distinction.

Is Amalia free to use?

Yes, Amalia is fully open source and free to use for most purposes, including commercial applications.

Which Portuguese variant does Amalia prioritize?

Amalia prioritizes European Portuguese (pt-PT), though it handles Brazilian Portuguese competently.

Where can I access Amalia?

Amalia is available on Hugging Face at amalia-llm/AMALIA-9B-0626-DPO. Managed hosting options are available through major cloud providers.

How does Amalia compare to other Portuguese LLMs?

On native pt-PT evaluations, Amalia substantially outperforms machine-translated alternatives and matches strong baselines on translated benchmarks, making it one of the top choices for European Portuguese applications.

Is there a roadmap for Amalia?

Yes, the consortium has announced plans for multimodal extensions and domain-specific fine-tunes for healthcare, law, and education.

Final Verdict

Amalia LLM represents a meaningful step forward for European Portuguese in the AI space. By prioritizing pt-PT as a first-class dimension and building on the proven EuroLLM-9B foundation, the model delivers authentic, culturally aware Portuguese that goes beyond surface-level translation.

For anyone working with European Portuguese—whether in content creation, business, research, or education—Amalia is a compelling choice. Its open-source nature, strong performance on native evaluations, and growing ecosystem make it a model worth watching in 2026 and beyond.

Rating: 4.5 out of 5

Disclosure: This review is based on publicly available information and the team’s technical report. Some pricing details may vary by provider.