NVIDIA expands open-model portfolio aimed at Europe’s “sovereign intelligence” across industries

NVIDIA on Oct. 21 outlined a broad expansion of its open-model ecosystem, positioning the move as a way for organizations in Europe and beyond to build and customize AI systems while keeping more control over where models run and what data they use. In a blog post, the company described the latest updates as “open models” that can be tailored with local datasets and deployed on infrastructure organizations control, rather than relying entirely on closed systems.

At the center of the expansion is a new set of NVIDIA Nemotron models for language and agent development. NVIDIA said Nemotron 3.5 Super VL and Nemotron 3.5 Bolt are designed to support agent workflows that can handle both text and images, helping developers orchestrate tasks, reason through complex problems, and complete specialized activities while aiming for lower cost. The company also introduced NVIDIA Nemotron Retriever models for embedding, reranking, and document parsing, roles that are commonly used to improve how an AI system retrieves relevant information from large document collections.

NVIDIA tied the Nemotron line to its Nemotron Coalition, describing it as a group that includes European partners Black Forest Labs, H-Company, and Mistral. The company also pointed to a range of enterprise and ecosystem users. It said Dassault Systèmes, Domyn, Dream, Palantir’s customers, Salesforce, SAP, Shopify, Synopsys, and WPP have been working with Nemotron models and datasets to create customized models and domain-specific agents. Professional services firms Accenture and EY were also listed as participants.

The blog post emphasized that customization is not limited to model architecture. NVIDIA said it is building a post-training ecosystem, listing partners such as AfterQuery, AgileRL, AI21, Applied Compute, Baseten, CoreWeave, Distil Labs, Google Cloud, and others. These companies, NVIDIA said, support enterprise efforts to tailor Nemotron for domain-specific workloads, suggesting that multiple layers of training, tuning, and deployment are intended to be modular rather than one-size-fits-all.

For structured data and decision support, NVIDIA introduced the NVIDIA Kumo family of open foundation models, libraries, and tools. NVIDIA described Kumo as part of its “Agent Knowledge Foundation” and said it is intended to shorten the time needed to make predictions and reduce the cost and complexity of building separate machine learning pipelines for each task. The company said Kumo can forecast outcomes, score fraud risk, and recommend actions based on structured business data, with an approach that allows teams to start from pretrained models and fine-tune using proprietary data when higher performance is critical.

NVIDIA also announced Kumo Structured, which it described as models for relational and tabular prediction, forecasting, and anomaly detection. It paired that with NVIDIA Kumo Engine, an open library stack for developing and deploying structured data foundation models. Early work from companies including Cognite, Coinbase, Databricks, DoorDash, Revolut, and Tubi was cited, and NVIDIA said Snowflake customers are using Kumo models to accelerate AI development in areas such as advertising, energy, financial services, and retail.

On robotics and “world models,” NVIDIA introduced or expanded the NVIDIA Cosmos family of open models for building systems that can understand and interact with the physical world. The company said it is adding new Cosmos post-training and optimization tools designed to improve world models using less compute, alongside simulation checkpoints meant to evaluate behavior in scenarios like interactive driving and robotic manipulation before real-world testing. NVIDIA also described Cosmos Data Factory, which it said includes blueprints and core agent skills to help developers curate, label, generate, and evaluate training data, including driving scenarios, product defects, image augmentation, and event generation. It cited adoption by Porsche with Voxel51, ANYbotics with Encord, Zalando and CEVA Logistics with Sereact, and Siemens.

Beyond language, structured prediction, and robotics, NVIDIA’s updates reached into healthcare and quantum computing. The company introduced BioNeMo RFD4-Proteina, an open protein design model developed with the Baker Lab at the University of Washington’s Institute for Protein Design. NVIDIA said the model generates protein sequences and 3D structures aimed at helping researchers design proteins that bind to difficult targets, assemble into complex structures, and bind to small molecules. The company claimed performance across more than 10 protein design tasks and said it generates designs about 13 times faster than RFdiffusion3 in a benchmark of 200 amino acid proteins.

Finally, NVIDIA described a new open model family for quantum computing called Ising, saying it provides a common, extensible foundation to speed progress toward useful quantum supercomputers. The company tied Ising use to tasks spanning real-time hardware calibration through error correction, including support for different qubit modalities and tailoring to individual hardware needs. It cited IQM Quantum Computers using Ising to calibrate its quantum processors, SEEQC combining NVIDIA’s NVQLink platform with Ising for low-latency real-time quantum error correction, and AWS making the Ising family available to researchers by integrating it into AWS Parallel Computing Service (AWS PCS).

NVIDIA did not provide details in the blog post about pricing, availability dates, or how these open-model releases will map to near-term revenue. The company also did not disclose any specific financial targets tied to Europe’s “sovereign intelligence” framing, nor did it quantify how quickly enterprises are expected to adopt the new models. As with many open-model ecosystems, the most meaningful proof points may show up later in documentation, developer usage, and performance comparisons under customer workloads rather than in the initial announcement.

Looking ahead, the key watch items are whether NVIDIA’s open-model releases translate into broader platform adoption across enterprise AI stacks, especially for structured prediction and agent deployments that involve proprietary data. For quantum and protein design, researchers will likely look for reproducibility, integration paths, and benchmarking results beyond the figures provided. NVIDIA’s emphasis on deploying on infrastructure customers control also suggests a competitive storyline centered on portability and governance rather than purely on model accuracy.