Portugal releases Amalia-9B, its own open-source LLM

Portugal has unveiled Amalia-9B, a 9 billion parameter large language model developed to highlight the country's potential in artificial intelligence. Released under the permissive Apache 2.0 license, the model aims to foster open research and development. Two versions are available on Hugging Face: one fine-tuned with Supervised Fine-Tuning (SFT) and another aligned via Direct Preference Optimization (DPO), both accessible under the 'amalia-llm' namespace. The accompanying paper on arXiv details the training methodology and architecture.

While specific coding benchmarks are not yet disclosed, Amalia-9B represents a significant step for Portugal's AI ecosystem. The choice of the Apache 2.0 license ensures broad commercial and research use, and the inclusion of both SFT and DPO adaptations allows users to choose the preferred alignment approach. This release joins a growing trend of smaller, efficient open models and underscores Europe's increasing investment in sovereign AI capabilities.