Villanova.AI has released Villanova2, a family of open language and vision-language models designed and trained in Italy for English, Italian, German, Spanish and French, with open weights and open data.
There are two backbones: a dense model with 2.4B parameters, and a sparse mixture of experts with 15.5B total and 3.9B active parameters. Both were pre-trained on 4.7 trillion tokens with a context of up to 131k tokens, and each comes as a base model, an instruction-tuned assistant and a vision-language model for document understanding. On Villanova.AI's evaluation over 85 benchmarks, both rank first among similarly sized European open models in all five languages.
The Villanova2 models cite Multi-SFT, Villanova.AI's collection of permissively licensed multilingual instruction data. Its authors include SAILab members Kamyar Zeinalipour and Andrea Zugarini.