Vitruvian-1 is much more than the Italian DeepSeek...

In the dynamic landscape of artificial intelligence development, the Italian startup Asc27 has made its entrance with a cutting-edge model: Vitruvian-1. Closely following DeepSeek, this innovation stands out as a benchmark for AI vision in the Italian and European market.

Unlike large language models (LLMs) and DeepSeek itself, Vitruvian-1 was designed to be a lean model, but with performance comparable to that of the most renowned solutions. CEO Nicola Grandis, as reported by Il Sole 24 Ore, emphasized that Vitruvian-1 offers an extremely competitive cost: just one euro per million tokens, compared to DeepSeek's 2.54 dollars, already known for its aggressive price that has alarmed the big tech companies. The resources used for development were also contained compared to competitors: the training of Vitruvian-1 cost only a few tens of thousands of euros, while for DeepSeek the investment was about six million.

Despite its small size, with only 14 billion parameters, Vitruvian-1 has achieved impressive results. It scored 95.5% on the MATH test, which evaluates mathematical problem-solving skills, and 90.2% on the MMLU, which measures linguistic comprehension and reasoning on multiple topics.

Thanks to a data-centric approach and an advanced architecture, with a special and rigorous filtering process of its 120 billion tokens, Asc27 has reduced the level of low-quality content and included reliable sources.

The result is a model characterized by low costs, high performance, and full compliance with European regulations (AI Act). Vitruvian-1 is therefore proposed as a benchmark and best practice for Italy and Europe. The continent has an environment rich in talent and skills in the field of artificial intelligence, but to realize this potential, a strategic vision and supranational collaboration are needed. With Vitruvian-1, Asc27 demonstrates that Europe has the capabilities to compete on par with global AI giants.