In recent years Italy has developed a number of artificial intelligence language models trained with particular attention to the Italian language. The topic involves universities, research centers and companies, both public and private.

Several projects share the use of the Leonardo supercomputer, managed by the Cineca consortium. The common goal is to offer tools suited to businesses, public administration and sensitive sectors.

Minerva and ChatMinerva

Minerva is described as the first family of large language models trained from scratch on Italian. It was developed by the Sapienza University of Rome, through the Sapienza NLP group, together with Cineca and within the FAIR project.

The model was built on a large set of open-access Italian and English data. After the launch of Minerva 7B, ChatMinerva arrived, a conversational assistant designed to access the web in real time and to understand texts, images and documents. Source: ansa.it

Modello Italia by iGenius

Modello Italia is a large language model developed by the company iGenius in collaboration with Cineca. It is presented as open source and trained from scratch on Italian-language data.

The foundation model has a 9-billion-parameter architecture and was trained on more than one trillion words, including public sources, synthetic data and partner content. The stated goal is to support companies and public administration, including areas such as healthcare, finance and national security. Source: milanofinanza.it

Almawave’s Velvet and the Overall Picture

Velvet is the family of models developed by Almawave, a company of the Almaviva group. The variants were trained on Cineca’s Leonardo supercomputer, with attention to language preservation and to Italian in particular.

According to the company, Velvet models aim for a balance between performance and limited resource use, so they can also be deployed on smaller infrastructure. They are described as multilingual, with support for Italian, English, Spanish, Portuguese, German and French. Source: digitalvoice.it

Overall, these projects show a growing commitment by Italian players to developing national language models. Common traits include training on Italian, the use of public computing infrastructure and a focus on applications for businesses and administrations.

The plurality of models reflects an ecosystem that combines academic research and industry. The main differences concern model size, licensing, multilingual capabilities and target application sectors.

Original article: repubblica.it