Skip to main content
Back to timeline
NVIDIA ResearchSource publication:

NVIDIA widens its open-model family at GTC Berlin, letting organizations customize Nemotron, Cosmos and BioNeMo with local data

Synopsis

At GTC Berlin, NVIDIA expanded its open-model family from language to robotics, biology and quantum computing, paired with agent tooling and a secure runtime. Developers tailoring applications to local languages and industries need open-weight foundations they can deploy themselves. Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, Nemotron Retriever, Kumo, Cosmos 3.5, Alpamayo 2 Edge, BioNeMo and Ising form that portfolio, which organizations can customize with local data and run on infrastructure they control.

Source-provided article image: NVIDIA GTC Berlin: What’s Next in AI | NVIDIA Blog
NVIDIA Research

Interpretation

NVIDIA widened its open-model family from language to language, robotics, biology and quantum computing, and released it alongside agent tooling and a secure runtime. The portfolio targets applications tailored to local languages and industries, putting customizable open models and deployment tools into one release. The announced lineup includes Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, Nemotron Retriever models for embedding, reranking and document parsing, the Kumo family for structured business data, Cosmos 3.5 world models, Alpamayo 2 Edge for autonomous vehicles, BioNeMo for healthcare and scientific discovery, and Ising for quantum computing.

The accompanying tools let agents work inside bounded permissions and draw on company data to act, not just answer. An agent is only as useful as what it knows and only as trustworthy as the boundaries it runs within; the expanded Agent Toolkit and OpenShell reference architecture target exactly those two points. The announcement says enterprises use the open source NVIDIA Agent Toolkit to customize and control agents, ground their answers and actions in company data, and monitor and govern how they operate; the OpenShell reference architecture is described as part of the portfolio for safely deploying agents at scale.

European cloud and local hardware partners carry these models and tools onto regional compute, so customization and deployment can stay on locally controlled infrastructure. Regional developers need compute close to their data, open models they can customize, and full-stack infrastructure that is secure by design; the release presents these as one regional package. The announcement cites NVIDIA cloud partners across Europe, the Nordics, Africa and Central Asia expanding regional compute for training and inference and setting boundaries on what agents can access and do; RTX Spark targets personal agents on Windows PCs, DGX Spark targets developers on Linux, and DGX Station targets enterprise frontier workloads.

Perspective

This report covers the open models, agent tooling and regional deployment arrangements NVIDIA announced at GTC Berlin, for researchers and engineers interested in customizable models and local deployment; on comparative evaluation for specific tasks, the announcement mentions only that the Cosmos platform ranks in the top five on six open-model leaderboards and gives no comparative results for the other models on specific tasks.

The announcement is mainly capability descriptions and partner plans; apart from the Cosmos platform ranking in the top five on six open-model leaderboards, it reports no quantitative performance of these models on specific tasks and does not say how post-customization results would be measured.