NVIDIA Expands Open AI Models for Europe’s Sovereign Intelligence
NVIDIA expands open AI models for language, business data, robotics, driving, protein design and quantum computing, with controlled deployment in Europe.
Summary
On Oct. 21, 2026, NVIDIA expanded open models organizations can customize with local data and run on controlled infrastructure. New releases are Nemotron 3.5 Super VL and Bolt, Nemotron Retriever for embedding, reranking and parsing, Kumo Structured and Engine for relational, tabular, forecasting, anomaly and fraud tasks, Cosmos 3.5 and Data Factory for simulation and training data, Alpamayo 2 Edge for DRIVE AGX Thor vehicles, BioNeMo RFD4-Proteina for proteins, and Ising for quantum calibration and error correction. Cosmos adds lower-compute post-training and simulation checkpoints.
Nemotron coalition partners Black Forest Labs, H-Company and Mistral contribute data, research and expertise; users include Dassault Systèmes, Domyn, Dream, Palantir’s customers, Salesforce, SAP, Shopify, Synopsys, WPP, Accenture and EY. Post-training support comes from AfterQuery, AgileRL, AI21, Alice, Applied Compute, Baseten, CoreWeave, Distil Labs, EON, Fastino, Fireworks, Google Cloud, Mercor, Multiverse Computing, Nebius, Handshake AI, Plurai, Prime Intellect, River.AI, Scale AI, Trajectory and TNG. Kumo users include Cognite, Coinbase, Databricks, DoorDash, Revolut, Tubi and Snowflake customers; Cosmos Data Factory users include Porsche with Voxel51, ANYbotics with Encord, Zalando and CEVA Logistics with Sereact, and Siemens.
RFD4-Proteina, created with the Baker Lab at the University of Washington’s Institute for Protein Design, leads more than 10 tasks and ran about 13x faster than RFdiffusion3 on 200 amino acid proteins. IQM Quantum Computers uses Ising for calibration; SEEQC combines it with NVQLink and Digital Link; AWS integrates it into AWS PCS. Super VL is available through Hugging Face, ModelScope, OpenRouter, build.nvidia.com and partners; Bolt E4B 4GB is expected Friday, Oct. 23. Retriever, Kumo, Cosmos, RFD4-Proteina and Ising are available through listed platforms; Alpamayo arrives on GitHub and Hugging Face later in October.
Positives
- RFD4-Proteina led more than 10 protein-design tasks and generated designs about 13x faster than RFdiffusion3 on 200 amino acid proteins.
- Kumo lets developers predict from structured business data without building a separate machine learning pipeline for each task.
- Alpamayo 2 Edge brings real-time scene understanding and motion planning to DRIVE AGX Thor with lower computing requirements.
- Cosmos Data Factory coordinates data curation, labeling, generation and evaluation for driving, defects, image augmentation and event generation.
- AWS PCS integration lets researchers build with Ising before deploying beside latency-sensitive quantum processors.
- NVIDIA’s open models can be customized with local data and deployed on infrastructure controlled by the adopting organization.
Risks & concerns
- Nemotron 3.5 Bolt E4B 4GB remains scheduled rather than available, with release expected Friday, Oct. 23.
- Alpamayo 2 Edge will not reach GitHub and Hugging Face until later in October.
- Kumo may require proprietary-data fine-tuning when higher performance is critical.
- Cosmos evaluates driving and robotic behavior in simulation before real-world testing, leaving physical validation still necessary.
- Quantum decoding typically needs classical hardware colocated with the processor because latency constraints remain.