Nvidia woos wary Europeans with more open weights models

Nvidia has released a new set of open weights models at GTC Berlin, aiming to meet European demand for sovereign AI and give organisations more options to build AI systems using their own data. The releases include tools for search and document processing, as well as a compact autonomous driving model.

The Nemotron 3.5 embedding and reranking models handle images and text in 34 languages and range from one to eight billion parameters; Nvidia says its largest versions lead dense models on the ViDoRe V3.1 benchmarks. Its one-billion-parameter PDF parsing model is claimed to process about 10.5 pages per second on one H100. Nvidia also says it provides training data and recipes for many models, allowing customers to inspect how they were built and train customised versions.