Villanova2: open European language models built in Italy
12, November, 2026
Villanova.AI has published Villanova2, its second generation of generative AI models. The company was founded to lead the Villanova project. Both the model weights and the training data are open, so researchers, companies and public bodies can inspect the models, adapt them and deploy them on their own infrastructure. For organisations that cannot send sensitive information to third-party services, that is a practical requirement.
The release has two backbones. Villanova2-15B-A3B is a sparse Mixture-of-Experts model with 15.5 billion parameters in total, of which only 3.9 billion are active for any given input. This keeps computing costs close to those of a much smaller model. Its hybrid architecture combines Gated DeltaNet layers with conventional attention, which makes very long documents considerably less memory-intensive to process. Villanova2-2B is a compact dense model with 2.4 billion parameters. Both were pre-trained on 4.7 trillion tokens, handle up to 131,000 tokens of context, and work in English, Italian, German, Spanish and French.
Each backbone comes in three versions. There is a base model, an instruction-tuned assistant with tool calling and optional reasoning, and a vision-language model built for document understanding. The team has also published GGUF builds for llama.cpp, so the models can run on local machines and edge devices as well as in the cloud. This fits the cloud-edge continuum that IPCEI-CIS is designed to support.
How the models perform
Villanova.AI evaluated the models on 85 benchmarks across ten categories and all five languages. By the team’s own assessment, both backbones rank first among European open models of similar size in every supported language. The larger model is also competitive with much bigger systems, scoring 0.4 points behind Qwen3-8B and 2.4 points behind Qwen3-30B-A3B, which has twice as many parameters in total. Translation stands out as a particular strength: the models placed first among all models evaluated.
Document understanding is a central focus of the release. On MDU, a multilingual benchmark that Villanova.AI built from real-world PDFs, both vision-language models rank first in their size classes. Much of the work in public administration, law and industry still depends on reading complex documents, so this capability matters well beyond research settings.
Why it matters
Most capable open models today are developed outside Europe and trained mainly on English. Villanova2 shows that a European team can build efficient, openly released models that perform well across several EU languages and can run on European infrastructure, from data centres to devices at the edge. A full technical report is due shortly, and the models are available now on Hugging Face.