NVIDIA Launches Nemotron for Science and Open Chemistry Dataset

What happened
NVIDIA released Nemotron for Science on November 16, 2026. This package combines the Nemotron 3.5 Super VL model with LitmusBench v0.2, a chemistry dataset containing nearly 50,000 questions. The system is designed for training and evaluating AI agents in scientific tasks, supporting text, image, and video inputs. The model and dataset are available on Hugging Face and can be run locally on NVIDIA's DGX Spark system.
Key details
- Nemotron 3.5 Super VL has 120 billion total parameters and a context window of up to one million tokens.
- LitmusBench v0.2 includes nearly 50,000 chemistry questions and is licensed under CC-BY-SA-4.0 and OpenMDW-1.1.
- The system supports reinforcement learning with verifiable rewards using experimental data as automated verification.
- NVIDIA reports a 12% accuracy improvement from supervised fine-tuning and an additional 17% from reinforcement learning in drug metabolism prediction tasks.
- Research organizations like Argonne National Laboratory and Jülich Supercomputing Centre are integrating Nemotron into their scientific workflows.
Why it matters
Scientific researchers and organizations can now train and deploy AI agents for specialized chemistry and life sciences tasks using open models and datasets. This enables them to run experiments and workflows on their own infrastructure, reducing reliance on external cloud services and potentially lowering costs. The verifiable training approach allows researchers to measure AI performance directly against experimental outcomes, supporting more reliable automation in scientific research.
