Open SourceMedium importance
European project Villanova.AI opens Villanova2, a 15.5-billion-parameter MoE model with documented weights and training data, under Apache 2.0
The Villanova.AI project, funded by the European Union’s IPCEI-CIS program, published the Villanova2 family on Hugging Face. The main one is Villanova2-15B-A3B-Base, a mixture-of-experts model with 15.5 billion parameters, 3.9 billion active per token, trained from scratch on 4.7 trillion tokens in English, Italian, German, Spanish and French, with a 131,072-token window. The weights are released under the Apache 2.0 license and every training data source is documented. Averaged across 39 tests, it scored 73.21; the authors say it is the strongest European MoE model, but it still trails Qwen3-14B-Base (75.56) and Gemma-4-26B-A4B (76.21), according to Tokenstrom’s summary. The family includes chat and vision versions and a 2.4-billion-parameter model.
Published in edition no. 33 of Noticiário do Joãozinho ·
Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese
Original source
Hugging Face (VillanovaAI)Tokenstrom summary
https://huggingface.co/VillanovaAI/Villanova2-15B-A3B-Base