NVIDIA adds Kumo structured-data models and Nemotron Retriever to an open portfolio aimed at European sovereign AI
The AMW Read
NVIDIA's open-weight push now includes structured-data and retrieval models and enlists Mistral and European partners under a sovereign-AI pitch, which changes the competitive map for open models and data infrastructure.
NVIDIA adds Kumo structured-data models and Nemotron Retriever to an open portfolio aimed at European sovereign AI
On October 21, 2026, NVIDIA announced a larger family of open models covering language, robotics, biology and quantum computing, and pitched it to European organizations that want to customize models on local data and run them on infrastructure they control. The lineup includes Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, which shipped on October 15. It also adds new Nemotron Retriever models for embedding, reranking and document parsing, and a new Kumo family of open-weight foundation models and libraries for relational and tabular prediction, forecasting and anomaly detection. Cosmos 3.5 world models, Alpamayo 2 Edge for autonomous vehicles, BioNeMo and Ising round out the set. NVIDIA says European partners Black Forest Labs, H-Company and Mistral contribute to the Nemotron Coalition, and that SAP, Dassault Systèmes, Salesforce, Accenture and EY are building on Nemotron.
The real news is how far the offering now reaches beyond language models. NVIDIA is handing out open weights, datasets and training recipes in areas where many startups sell proprietary models: retrieval, tabular ML and domain fine-tuning. Mistral is a member of the coalition, which puts Europe's flagship model lab inside NVIDIA's program rather than in competition with it. The sovereign pitch also lines NVIDIA up with European governments and enterprises that want models running on local hardware. Most of that hardware will run on NVIDIA chips.
For builders, the pretrained Kumo models put pressure on vendors selling per-task ML pipelines for fraud scoring and forecasting. The same goes for retrieval-stack startups now competing with free Nemotron Retriever models. Post-training partners such as Baseten, Fireworks, Nebius and Prime Intellect are left competing on customization services, not model ownership. Performance claims still need independent reproduction.
