Nvidia courts European sovereign-AI buyers with refreshed Nemotron open models at GTC Berlin

Nvidia used its GTC Berlin keynote to deepen a pitch aimed squarely at European governments and enterprises wary of dependence on US and Chinese closed models: build on open weights, keep the data at home, and run it on Nvidia silicon. The centrepiece was a refresh of the Nemotron Retriever family, rebuilt on Nemotron 3.5, which embeds both text and images across 34 languages in sizes from one billion to eight billion parameters, with NVFP4-quantised variants for Blackwell accelerators.

Nvidia claims the eight-billion-parameter Retriever models are the top-ranked dense models on the ViDoRe V3.1 visual document retrieval benchmark at BF16 precision. Alongside them came a refreshed Nemotron Parse, trained to pull structured information out of PDFs, which the company says processes around 10.5 pages per second on a single H100 in its one-billion-parameter form, and Alpamayo 2 Edge, a compact autonomous-driving model built for real-time video and sensor processing on the DRIVE AGX Thor platform.

The Berlin announcements bookend a busy fortnight. A week earlier Nvidia launched Nemotron 3.5 Super VL, a vision-language model, and teased Nemotron 3.5 Bolt, a 74-billion-parameter mixture-of-experts model designed to fit in as little as 24 GB of video memory. The Register reported that Bolt would be downloadable from the Friday after the keynote.

What distinguishes Nvidia's approach from most open-weight releases is that it routinely publishes training data and recipes with its models. That lets customers post-train on the Nemotron architecture with their own corpora rather than merely fine-tuning, an approach the Allen Institute's OLMo project has also pursued. Nvidia's bet is that most enterprises lack the resources to train from scratch and will instead customise open models from vendors such as Mistral or H Company, all of it running on its GPUs.

Why it matters

Retrieval and document parsing are the unglamorous layers that decide whether enterprise RAG systems work. Multilingual, image-aware embedders released free of charge, with quantised builds for the latest hardware, lower the cost of building serious search and knowledge systems without sending documents to a third-party API.

Indicops view

Indian government and enterprise programmes face the same sovereignty questions Nvidia is answering for Europe. A 34-language embedder and a fast PDF parser that run on-premise are directly useful for departmental document search, tender analysis and case-file retrieval, where data cannot leave the perimeter. We would evaluate the Nemotron Retriever models against existing Indic embedding options on real departmental documents, since benchmark leadership in European languages does not guarantee the same in Hindi, Tamil or Marathi.