Velvet initiative aims to build a Multimodal Large Language Model to increase the European capacity and sovereignty in the field of Artificial Intelligence. In particular, Velvet models will augment the successful Velvet family of LLMs, both by exploiting a Mixture-of-Expert (MoE) architecture and adding further capabilities.
Currently Velvet is a family of large language models, developed from scratch, featuring a dense architecture. The models have been trained on the EuroHPC Leonardo infrastructure hosted by CINECA, utilizing public data that underwent extensive curation with a strong emphasis on data governance, alignment with EU standards and a broad work on ethical AI since its design phase.
In addition to all European Languages, Velvet will also support some African languages, thus opening European LLMs to the important market of the African languages.
Raniero Romagnoli, Almawave SpA, Italy