BSC-LT/ALIA-40b-instruct is a 40B instruction-tuned LLM from the Barcelona Supercomputing Center’s Language Technologies group. It’s trained from scratch and positioned as a broadly multilingual assistant: the base model covers 35 European languages (plus code), and the post-training focuses most heavily on Spanish, Catalan, Basque, Galician, and English.
The most notable spec is the context window: the model card lists a context length of 163,840 tokens after long-context adaptation. That makes it a compelling candidate for “long document” workflows (multilingual retrieval, translation + summarization, or Q&A over big reference material), assuming you have the hardware to run a 40B checkpoint.
Two practical notes from the authors: (1) the model is only preliminarily aligned (they explicitly warn about biased/incorrect/harmful output), so treat it as a research/development model unless you’re adding your own safety layer; and (2) this checkpoint is marked as superseded by a newer ALIA-40b-instruct-2512 release, so you’ll likely want to start there if you’re evaluating ALIA today.
Quick stats from the listing feed: pipeline: text-generation · 1 like · 280 downloads.
Source listing: https://huggingface.co/models?sort=modified