AMÁLIA and the future of European Portuguese LLMs

AMÁLIA is a €5.5M Portuguese government-funded open source LLM for European Portuguese, built as a continuation of EuroLLM's pre-training rather than trained from scratch. The author analyzes the technical report and raises concerns: the model weights, data, and benchmarks are not yet publicly available despite open source claims; only ~5.5% of pre-training tokens are clearly European Portuguese; and the new benchmarks don't measure intrinsic knowledge about Portugal. Despite beating models like Qwen 3-8B on most Portuguese benchmarks, the author questions whether the approach is fully optimized and calls for more Portuguese data, genuine openness of weights and datasets, and better evaluation dimensions.