Nvidia woos wary Europeans with more open weights models

Nvidia is targeting European enterprises seeking sovereign AI by introducing a range of new open weights models for applications in chatbots, healthcare, and robotics. The tech giant unveiled its latest additions to the open weights model library at GTC Berlin, including refreshed Nemotron Retriever embedding and Alpamayo autonomous vehicle models.

Embedding models, while not the most glamorous, play a crucial role in retrieval augmented generation (RAG) technology, which Nvidia's new models support for both images and text in 34 languages. These models, based on Nemotron 3.5, come in sizes from one to eight billion parameters with NVFP4-quantized versions for Blackwell-based accelerators.

Nvidia claims its largest model is currently the top-ranked dense model on the ViDoRe V3.1 benchmarks. To assist with converting documents into embeddings, Nvidia has also launched a refreshed Nemotron Parse model that can process approximately 10.5 pages per second on a single H100. Additionally, Nvidia is providing access to Alpamayo 2 Edge, a more compact autonomous driving model that can process streaming video and sensor data in real-time on the DRIVE AGX Thor platform.

These new models complement Nvidia's previous releases, such as the Nemotron 3.5 Super vision language (VL) model and the recently teased Nemotron 3.5 Bolt, a 74 billion-parameter mixture of experts (MoE) model. Unlike many open weights models, Nvidia makes available the training data and the processes used to create its models, allowing customers to inspect the models and build custom versions based on the Nemotron architecture.