NVIDIA has unveiled “Nemotron for Science,” a newly integrated suite of AI tools and resources aimed at accelerating scientific breakthroughs, according to a November 16, 2026 announcement. This initiative combines cutting-edge generative models, expansive chemistry datasets, and a toolkit for customizing AI agents that can reason through complex scientific problems.

Central to the offering is the Nemotron 3.5 Super VL model, a large language model designed specifically for scientific tasks and research workflows. By leveraging post-training recipes tailored for life science and chemistry, Nemotron enables researchers to adapt the system for highly specialized purposes — spanning from laboratory automation to the rational design of proteins and novel materials.

The platform introduces LitmusBench v0.2, an open-access chemistry dataset boasting nearly 50,000 questions. This resource serves the dual purpose of training and objectively evaluating language models for their capacity to perform chemistry reasoning, make accurate predictions about molecular behavior, and interpret experimental data. NVIDIA says these tools give scientists the power to create agents capable of extracting chemical knowledge, planning complex experiments, and generating protocols for laboratory use.

Collaborations with prominent research organizations are already underway. The U.S. Department of Energy’s Argonne National Laboratory is integrating Nemotron into the StructBioReasoner platform to design and assess therapeutic proteins, while Sweden’s KTH Royal Institute of Technology is tapping the system for autonomous repairs of molecular dynamics setups. Japan’s RIKEN research institute will provide broader access to Nemotron for Science via its AI supercomputing platform.

Real-world scientific applications are emerging quickly. At Terray Therapeutics, the model enabled precise extraction of chemical reactions from scientific diagrams, outperforming previous benchmarks and speeding the conversion of published research into actionable laboratory data. Lila Sciences used post-trained Nemotron tools for designing quantum dots — nanoscale materials key to next-generation screens and solar panels — achieving higher accuracy at a lower compute cost than the latest commercial frontier models.

For researchers interested in deploying these capabilities, Nemotron 3.5 Super VL and the LitmusBench v0.2 dataset are now available on the open AI repository Hugging Face. Furthermore, the platform is engineered to run efficiently on 128GB NVIDIA DGX Spark workstations, offering scalable solution from desktop experimentation to supercomputing deployments.

This launch marks NVIDIA’s latest step in democratizing advanced AI for science, building on the foundation laid by the Nemotron 3 family of models introduced in late 2025. By distributing open models and domain-specific tools, NVIDIA aims to increase the pace of discovery across a spectrum of scientific disciplines. Specific details on hardware pricing and acquisition pathways were not disclosed in the initial announcement.