Nvidia expands its open models for RAG use cases
At GTC Berlin, Nvidia presented new open models for research, health, and robotics. The company aims to attract European businesses and administrations seeking more inspectable AI components.
At GTC Berlin, Nvidia unveiled a new series of open-weight models for European businesses and administrations, with a clear focus on RAG workflows, healthcare, robotics, and autonomous driving. The goal is clear: to offer building blocks that are easier to inspect and adapt than closed models, as the demand for AI sovereignty grows in Europe. For teams that lack the resources or time to train everything themselves, the message is simple: there are now more modular options, not just polished black boxes.
More open models, but not completely opaque
Nvidia relies on a fairly broad definition of open weights: the model weights are downloadable, but the company goes further than many competitors by also publishing the training data and the recipes used to obtain the final model. This approach changes the game for clients who want to audit, reproduce, or rebuild a variant on their own Nemotron architecture, rather than settling for simple fine-tuning afterwards. The Allen Institute follows a similar logic with its OLMo models.
The contrast is stark with most so-called open models, often distributed without detailed explanations of their creation. In this context, Nvidia tries to convince European buyers that useful openness is not limited to downloading. It is also a matter of method, traceability, and technical control.
RAG, PDFs and videos: what the new models can do
The new building blocks presented cover several stages of a RAG pipeline, that is, a method that splits documents, converts them into vectors, then ranks them before injecting them into the context of an LLM. Nvidia launched new embedding and reranking models based on Nemotron 3.5, capable of processing text and images in 34 languages. They come in sizes ranging from 1 to 8 billion parameters, with NVFP4 quantized versions for Blackwell accelerators.
The company claims that its 8-billion-parameter model now tops density rankings on the ViDoRe V3.1 benchmark in BF16 accuracy [to be verified]. Nvidia also introduced Nemotron Parse, a model trained to extract information from PDF documents, which are often difficult to exploit because they mix text, tables, and scans. According to the company, the 1-billion-parameter version processes about 10.5 pages per second on a single H100.
In the same batch, Nvidia made available Alpamayo 2 Edge, a more compact version of its autonomous driving model, designed for real-time processing of video streams and sensor data on the DRIVE AGX Thor platform. A week earlier, the company had already showcased Nemotron 3.5 Super and teased Nemotron 3.5 Bolt, a 74-billion-parameter MoE announced to fit within 24 GB of video memory. Nvidia promises a download starting Friday.
Why this strategy appeals to European buyers
The subtext is as commercial as it is technical. In Europe, the demand for so-called sovereign AI solutions pushes organizations to prefer models they can host, adapt, and document. Nvidia is targeting this segment precisely with ready-to-use components that are open enough to reassure security, compliance, and data teams.
The strategy remains pragmatic: few companies have the capacity to build a large model from scratch, while an open and well-documented foundation allows accelerating internal projects without starting from zero. In practice, this makes AI less spectacular but more usable. And that is often where the budget is decided.
Key points
- Nvidia presented these models at GTC Berlin.
- The new embedding models support 34 languages.
- Nemotron Parse processes about 10.5 pages per second.
- Alpamayo 2 Edge targets real-time on DRIVE AGX Thor.
- Nemotron 3.5 Bolt is designed to fit within 24 GB of video memory.
- The Allen Institute follows a similar approach with OLMo.
By the numbers
- 34 languages — supported by the new embedding and reranking models, according to Nvidia.
- 8 billion parameters — maximum size announced for the new embedding and reranking models.
- 10.5 pages per second — throughput announced for Nemotron Parse on an H100.
- 74 billion parameters — size of Nemotron 3.5 Bolt, according to Nvidia.
- 24 GB — video memory targeted for Nemotron 3.5 Bolt.