NVIDIA's open model family expanded at GTC Berlin on October 21, 2026. In its official blog post published that day, NVIDIA added open models spanning language, robotics, biology and quantum computing, which organizations can customize with local data and deploy on infrastructure they control — exactly what Europe's sovereign intelligence debate has been missing.
What was added
On language, Nemotron 3.5 Super VL and Nemotron 3.5 Bolt arrive alongside Nemotron Retriever for embedding, reranking and document parsing. For structured data, the Kumo open-weight foundation models help agents forecast outcomes, score fraud risk and recommend actions from business tables. For the physical world, Cosmos 3.5 world models and Alpamayo 2 Edge bring real-time reasoning into vehicles. BioNeMo RFD4-Proteina, built with the University of Washington's Baker Lab, generates designs about 13x faster than RFdiffusion3 on a 200-amino-acid benchmark. The Ising family serves quantum processor calibration and real-time error correction.
Why European enterprises care
Sovereign intelligence is not about another cloud API; it is about controlling weights, data and deployment location. Dassault Systèmes, SAP, Salesforce and Shopify are already customizing Nemotron, while Cognite, Coinbase, Databricks and DoorDash are early Kumo users. Most new models are on Hugging Face and offered as NIM microservices, with Bolt E4B expected on October 23. The release lowers the barrier to entry, but post-training data, compliance review and local compute remain the buyer's responsibility — open weights do not mean compliance out of the box.
