ChatMinerva and Minerva AI LLM: the Italian Large Language Model family pretrained from scratch

The first family of LLMs pretrained from scratch on Italian!

The First Italian-centric Large Language Model

Minerva is the first family of large language models pretrained from scratch in Italian, built on trillions of words from open-access Italian and English data. Initially developed by the Sapienza NLP group led by Professor Roberto Navigli, in collaboration with FAIR and powered by CINECA’s Leonardo supercomputer, it represents a new standard for high-performance, sovereign AI.

Designed as a truly open model in both data and architecture, Minerva combines strong linguistic alignment with scalable infrastructure.

Now entering its next phase, Minerva is being further developed and industrialized by Babelscape into a robust, production-ready solution for sectors including Public Administration, healthcare, defense, culture, and education.

Try ChatMinerva now

Public University

Developed by Sapienza University of Rome

Truly Open-Source

Models and training data are fully accessible

Constant Maintenance

Updated by Sapienza and Babelscape

Safety-Aware

Designed to prevent misinformation

Deployed in Public Sector

Trusted by public administration sectors

Controlled Training Data

Transparent and curated sources only

Introducing ChatMinerva

ChatMinerva radically expands the capabilities of the Minerva LLM, transforming it into a true multimodal AI assistant.
Users can now upload photographs, images, scanned pages, scientific papers, reports and technical documentation, asking ChatMinerva to interpret, summarize, analyze or answer questions about their content.
As the first large language model pretrained in Italian, Minerva has always stood for technological independence. With ChatMinerva, that commitment takes a concrete form: a capable, reliable AI experience built entirely on Italian ground. A "zero-mile" model designed to serve both the public and private sectors.

Multimodal

Upload pictures, photos, scanned pages and ask Minerva about them.

Live web access

Minerva now answers using real-time knowledge from the Web.

Long document handling

Attach papers, reports and documentation and chat about it.

Enhanced safety

A specific component ensures that outputs are controlled and trustworthy.

Our Minerva models

Base model

Our pretrained language models span from 350 million to 7 billion parameters, offering scalable solutions for various needs.

The smaller models are fast and lightweight, ideal for quick responses and low-resource environments, while the mid-sized options balance depth and efficiency for more complex tasks. All models are dual-trained in Italian and English, making them versatile for both general-purpose and fine-tuned domain-specific applications.

Instruction tuned

The instruction-tuned version of the 7 billion-parameter model is specifically adapted for tasks such as conversation, summarization, machine translation, and other language processing applications.

This tuning enhances its ability to follow user instructions while incorporating safeguards to improve the quality and safety of its outputs. Its design makes it suitable for a variety of use cases requiring contextual understanding and task-specific adaptability.