ALIA 40B Instruct 2601

ALIA-40b-instruct is the instruction-tuned release of Spain's public language model, built by the Language Technologies Lab at the Barcelona Supercomputing Center and trained from scratch on MareNostrum 5, the EuroHPC pre-exascale system BSC hosts. The base model saw 9.83 trillion tokens of curated data across 35 European languages and code, and the 40.4-billion-parameter architecture runs 48 layers with grouped-query attention. The context-extended base supports up to 163,840 tokens, though the window an individual route serves may be shorter. Post-training covers 808k instruction conversations and 368k preference pairs, concentrated on Spanish, Catalan, Basque, Galician, and English. The 2601 checkpoint improves on 2512 with stronger instruction adherence, better resilience to noisy or malformed input, tightened safety alignment, and improved long-context behaviour. Conversations use the ChatML template. Weights, training scripts, and configurations are released under Apache 2.0, with funding from Spain's Ministry for Digital Transformation and support from the Catalan government's Aina project. It fits multilingual assistants, summarisation, and translation work across Iberian and other European languages, particularly where a fully open, publicly funded model is preferred over a commercial one.