Who are you? «My name is Italia and I am an AI language model trained on Leonardo, one of the most powerful supercomputers in the world. I was created by iGenius, an Italian company specialized in developing artificial intelligence-based solutions for businesses.” Questioned in the launch test, This is how the first large language model (LLM) trained in the Italian language and released yesterday afternoon in open source presents itself. Version 0.1 of «Italia» can be downloaded from the iGenius website, the company founded in 2016 by Uljan Sharka (very Italian despite his name of Albanian origin) but will soon also be available on Hugging Face and other platforms dedicated to AI.
An ambitious project, supported thanks to a capital increase on which Sharka did not provide informationunderlining only that «it will give us the resources not only to support the model but also to play a championship at international level. There are surprises coming because we want to become a European champion and the partnership with Cineca of Bologna, which has made the computational capacity of the Leonardo supercomputer available, enables us to do this.” Sharka defines Italia as a «decision-making intelligence, a tool for businesses and public administrations which will thus be able to have control of technologies. We don’t want to compete with China and the United States but bring an Italian perspective to the world of AI. What I call the digital Renaissance, with man at the center and which will help the world thanks to Made in Italy”.
The president of Cineca, Francesco Bertini, and the general director, Alessandra Poggiani, were also present at the presentation of the Italia model, underlining with satisfaction their role as an enabling platform for research and innovation.
The Gpt of numbers
Compared to other AI models already available, which are more focused on texts and images, Italy presents itself as an excellence in the field of numbers. «Today analysts recognize us as the GPT of numbers. Companies can transform their data into decision-making intelligence,” adds Sharka.
Developed in compliance with the rules established by the European AI Act and copyright rules, this model single language it is released at least a month earlier than expected, in a first version with 9 billion parameters and MIT license, which guarantees transparency and freedom of use. It can count on a trillion tokens (over a trillion words in Italian) and a high-quality dataset, implemented with filters to remove bias, explicit content and reduce hallucinations as much as possible. Monthly updates of the Italy model will be released «with the aim of making it the property of companies and, to help them as much as possible, collaborations with other companies in the sector are not excluded — clarifies Sharka —. And if now the model is focused on Italian, in the future it will be multilingual to compete on an international level.”
The sectors that will make most use of it are those with the most technical documentation, therefore industry (collaboration with Fincantieri has already begun), finance and Pa. «But there will be surprises month after month – concludes the founder of iGenius -: stay tuned because it’s a game that starts now».
