It's not just a name evoking the tradition of fado: Amalia is the first Large Language Model developed under the aegis of the Portuguese government, a project that marks a concrete step in the country's digital sovereignty strategy. With 9 billion parameters, an Apache 2.0 license, and two variants already available on Hugging Face—one with supervised fine-tuning (SFT) and the other optimized via Direct Preference Optimization (DPO)—the model joins a European landscape where the race for national LLMs is accelerating.

The release doesn't come with specific coding benchmarks, a detail that may weigh on evaluation teams. But Amalia's real value lies elsewhere: in its architecture designed for Portuguese, in a training dataset and language preference ecosystem that global models—no matter how massive—struggle to cover with the same depth. It's no coincidence that Portugal chose to invest in a local project: handling sensitive data, regulatory compliance, and the need for fast inference on domain-specific tasks (from public administration to legal sectors) push toward controllable solutions.

Why the license choice matters

Apache 2.0 is more than a technical footnote: it enables unrestricted on-premise deployment, code modification, and commercial distribution. For a public body or a company that must keep data within national borders, being able to run the model on proprietary infrastructure—without depending on third-party cloud APIs—is a decisive factor. The 9 billion parameter size, moreover, configures an accessible hardware profile: with quantization techniques, it's realistic to consider running it on GPUs with VRAM in the tens-of-gigabytes range, reducing TCO compared to larger models.

Missing benchmarks aren't an oversight

The absence of standardized coding performance metrics—flagged in the original discussion—reflects a priority choice: Amalia wasn't built to compete on generic tasks with hundred-billion-parameter models, but to become an infrastructural building block for Portugal's digitization. That doesn't absolve adopters from running internal tests on retrieval-augmented generation, summarization, and legal-administrative language understanding, areas where training on local corpora can make a real difference.

In the bigger picture, the Portuguese case confirms a trend: the proliferation of national LLMs isn't just about flag-waving, but a pragmatic response to cloud vendor lock-in and the need to align models with cultural and regulatory contexts. For those tracking on-premise architectures, Amalia is yet another sign that the market is gearing up to offer genuinely self-hosted alternatives, where data sovereignty is the norm rather than the exception.