At GTC Berlin, Nvidia expanded its open-weights model catalog with offerings aimed at European sovereign AI demand. The new releases include Nemotron Retriever embedding and reranking models, a refreshed Nemotron Parse model for PDFs, and Alpamayo 2 Edge, a compact autonomous-driving model for real-time sensor and video processing on the DRIVE AGX Thor platform.
The embedding models, based on Nemotron 3.5, support both text and images across 34 languages and come in sizes from 1B to 8B parameters, with NVFP4-quantized versions for Blackwell accelerators. Nvidia claims the 8B model is the top-ranked dense model on the ViDoRe V3.1 benchmark at BF16 precision. For document-heavy workflows, Nemotron Parse is trained specifically to extract content from PDFs, with a claimed throughput of roughly 10.5 pages per second when running the 1B-parameter variant on an H100.
Nvidia also teased Nemotron 3.5 Bolt, a 74B-parameter mixture-of-experts model designed to fit in as little as 24GB of video memory, with downloads promised for Friday. Unlike most open-weights releases, Nvidia publishes the training data and recipes used to build its models, allowing customers to create fully custom models rather than fine-tune black-box weights. The Register notes that the Allen Institute has taken a similar approach with its OLMo models.