Nvidia has expanded its open-model portfolio beyond language systems into structured enterprise data, robotics, autonomous vehicles, protein design and quantum computing. The releases are intended for organizations that want to customize models with local data and deploy them on controlled infrastructure.
New Nemotron 3.5 language and retrieval models support agents working with text, images and documents. The Kumo family handles tables and relational data for forecasting, anomaly detection and fraud scoring, reducing the need to build a separate machine-learning pipeline for every prediction task.
Cosmos 3.5 adds world models and data-generation tools for robotics and driving simulations. Alpamayo 2 Edge brings scene understanding and motion planning to Nvidia’s in-vehicle Drive hardware. BioNeMo RFD4-Proteina generates protein sequences and 3D structures, while the Ising family targets quantum calibration and error correction.
Most announced models are available through combinations of Hugging Face, GitHub and Nvidia’s NIM services. Nemotron 3.5 Bolt is expected October 23, and Alpamayo 2 Edge later this month. Nvidia reports strong performance claims, including 13-times-faster protein generation than RFdiffusion3, but those vendor benchmarks need workload-specific validation.