Nvidia Unveils Nemotron Retriever Models for Multilingual RAG at GTC Berlin

Nvidia introduced new open weights models at GTC Berlin to address European demand for sovereign artificial intelligence capabilities. According to The Register, these additions include refreshed Nemotron Retriever embedding and Alpamayo autonomous vehicle models. The move targets enterprises that lack the resources to build models from scratch and instead rely on fine-tuning existing open weights foundations.

What specific models were released?

Nvidia added to its open weights library with updated Nemotron Retriever embedding models and Alpamayo models for autonomous vehicles. The Register reports that the embedding models are based on Nemotron 3.5 architecture. They are designed to support both image and text embeddings across thirty-four languages. This expansion aims to provide more capable tools for enterprise applications that require multilingual processing and visual data integration.

How do these models fit into RAG workflows?

Retrieval augmented generation relies on converting documentation chunks into vector embeddings before ranking them for relevance. The Register explains that Nvidia’s latest models handle this embedding and reranking process. The new offerings include sizes ranging from one to eight billion parameters. Smaller sizes allow deployment on devices like the DGX Spark, leaving room for other concurrent model operations in enterprise environments.

What performance claims accompany these releases?

Nvidia states that its new eight-billion-parameter embedding and reranking models lead the ViDoRe V3.1 benchmarks for dense models at BF16 precision. The Register notes that vendors’ performance claims should be viewed with caution. However, the availability of NVFP4-quantized versions for Blackwell-based accelerators suggests an optimization path for newer hardware. This quantization allows for efficient processing on updated datacenter infrastructure.

What limitations remain for document processing?

Not all document types convert easily into usable embeddings for these models. The Register highlights that PDFs are notoriously difficult for embedding models to interpret correctly. This limitation persists despite the improved multilingual and multimodal capabilities of the new Nemotron-based models. Enterprises using heavy PDF documentation for their RAG pipelines may still encounter parsing challenges that require additional preprocessing steps.

Questions to ask your suppliers

Security teams should verify whether suppliers have assessed the provenance and integrity of the open weights models they integrate into production systems. Operations managers need to confirm that internal policies address the risks of using third-party foundation models, particularly regarding data leakage during fine-tuning. IT leaders must ensure that the hardware supporting these models meets security isolation requirements. Governance committees should review whether the chosen models align with organizational data residency and sovereignty mandates.

Background: Embeddings

Embeddings are numerical representations of words, sentences, or objects placed in a high-dimensional space. Software uses these coordinates to measure semantic relatedness, turning text and data into vectors that algorithms can compare through distance rather than exact keyword matches.

Read the full guide: Understanding embeddings and how machines measure meaning

What to do and how to stay safe: Nvidia

• Verify that the source of the open weights models is trusted and that the model files have not been tampered with before deployment.
• Monitor system logs for unusual behavior when processing new document types, especially PDFs, through your retrieval augmented generation pipelines.
• Ensure that your hardware configurations support the specific quantization formats provided by the vendor to maintain intended performance and security boundaries.
• Review your data handling procedures to confirm that sensitive information is not inadvertently exposed during the embedding or fine-tuning processes.

Step-by-step guide: AI Usage Policies at Work: How to Prevent Data Leakage

General security guidance from the TechSec Feed newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Which Nvidia models were unveiled at GTC Berlin?

Nvidia unveiled refreshed Nemotron Retriever embedding models and Alpamayo autonomous vehicle models.

How many languages do the new embedding models support?

The new Nemotron 3.5-based models support embeddings across thirty-four languages.

What is a known limitation of these embedding models?

PDFs are notoriously difficult for these embedding models to process effectively.