NVIDIA announced a wider open-model portfolio spanning language, structured business data, robotics, protein design and quantum computing on October 21, 2026.
NVIDIA announced a broad expansion of its open-model portfolio on October 21, 2026, with additions spanning language and vision, enterprise data, robotics, biology and quantum computing. The company says organizations can adapt the models with local data and deploy them on infrastructure they control.
The releases include Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, plus Nemotron Retriever models for embedding, reranking and document parsing. NVIDIA Kumo adds models and tools for structured business data, including forecasting, anomaly detection and prediction from relational and tabular information. NVIDIA says developers can fine-tune Kumo with proprietary data; Cognite, Coinbase, Databricks, DoorDash, Revolut and Tubi are among the companies working with the models.
For physical AI, Cosmos 3.5 includes models and tools for training robots, autonomous vehicles and vision agents. NVIDIA also introduced Cosmos Data Factory workflows for preparing and evaluating training data. Alpamayo 2 Edge is a compact reasoning model for real-time scene understanding and motion planning in vehicles, optimized for NVIDIA’s DRIVE AGX Thor platform.
The biology release, BioNeMo RFD4-Proteina, was developed with the University of Washington’s Baker Lab and generates protein sequences and 3D structures. NVIDIA reports that it performed across more than 10 protein-design tasks and generated designs about 13 times faster than RFdiffusion3 in a benchmark involving 200-amino-acid proteins. Ising, another new family, targets quantum-computing tasks such as processor calibration and error correction.
Availability varies by release. Nemotron 3.5 Super VL, Nemotron Retriever, Kumo, Cosmos, BioNeMo RFD4-Proteina and Ising are listed as available through specified model platforms or NVIDIA services. NVIDIA said Nemotron 3.5 Bolt E4B 4GB was expected October 23, while Alpamayo 2 Edge was expected later in October. The portfolio gives developers several distinct model families to evaluate for specialized applications, rather than a single general-purpose system.