National model programme, Italy
Rome
Sapienza University's Minerva and Almawave's Velvet: Italy's home-grown Italian-language models.
Rome is the centre of Italy's effort to build models trained for Italian. Sapienza University of Rome's NLP group released Minerva, a family trained from scratch on Italian and English, and the Rome-based company Almawave released the Velvet models. Both used Leonardo, the supercomputer run by the CINECA consortium near Bologna, one of Europe's most powerful.
Minerva and Velvet
Minerva, led by Roberto Navigli at Sapienza and first released in 2024, was built within the FAIR (Future Artificial Intelligence Research) national programme and trained with most of its data in Italian. Almawave's Velvet family, including Velvet 14B and Velvet 2B, are open-weights multilingual models with a focus on Italian, designed for energy efficiency and offered to Italian businesses and public bodies.
Italian policy
Italy passed a national AI law in September 2025, the first EU country to add its own law alongside the AI Act. Its data protection authority temporarily blocked ChatGPT in March and April 2023 over privacy concerns and later fined OpenAI, a case that made Italy an early test of how European law applies to chatbots. Public funds also back AI start-ups and computing.