A science-focused agent package
NVIDIA introduced Nemotron for Science, built on its open Nemotron 3.5 Super VL model. The goal is to help researchers build specialized agents for their own fields. The package combines the model with evaluation sets, an open chemistry dataset and post-training recipes.
LitmusBench v0.2 and chemistry data
At the center is LitmusBench v0.2, an open dataset of nearly 50,000 questions. Drawing on molecular properties and experimental drug discovery data, it is designed both to evaluate and to improve models' chemistry reasoning. Researchers can use it to boost a model's ability to predict how molecules behave, including solubility and how the body processes them.
Post-training and notable results
Researchers can post-train Nemotron models with domain data to specialize them. According to NVIDIA's examples, Terray Therapeutics trained the model to extract reactants, products and reaction conditions from chemistry diagrams, recovering about 76% of annotated reactions on the RxnScribe test. Lila Sciences raised accuracy for making quantum dots to 61% for its automated labs; the company says this beats GPT-6 Astra's 54% at 10-20x lower inference cost.



