Quick hits: March 2025

Almawave unveils Velvet: italy’s innovative approach to AI

Italian AI firm Almawave has introduced Velvet, a multilingual generative AI model designed and developed entirely in Italy. The announcement was made at Auditorium della Tecnica in Rome, marking a significant step in Italy’s AI development.

Velvet is a large language model (LLM) capable of processing and generating complex text across six languages—Italian, English, French, German, Portuguese, and Spanish—with plans to expand into Swahili through a partnership with Tanzania’s Information & Communication Technologies Commission.

The models, Velvet 14B and Velvet 2B, were trained using the Leonardo supercomputer, operated by Cineca, and will be released as open-source starting January 30. Primarily aimed at business applications, these models stem from over 15 years of R&D, integrating AI across industries such as healthcare, finance, mobility, public administration, and security. According to Almawave CEO Valeria Sandei, Velvet was designed to be energy-efficient, adaptable, and fully compliant with European AI regulations. Unlike other AI models, Velvet focuses on linguistic diversity, ensuring strong support for Italian and other European languages through proprietary training techniques.

A major innovation in Velvet is Privacy Association Editing (PAE), an advanced feature that enables the removal of sensitive data without requiring model retraining—a key breakthrough in AI privacy management.