NVIDIA has introduced Nemotron for Science, a package built on its Nemotron 3.5 Super VL open model that gives researchers the pieces to build specialised AI agents for scientific discovery: the model itself, a new open chemistry dataset, scientific evaluations and post-training recipes.

The dataset, LitmusBench v0.2, holds nearly 50,000 questions for training and evaluating language models on chemistry reasoning, drawing on molecular properties and experimental drug-discovery data to test predictions about how molecules behave, including their solubility and how the human body processes them. Post-training lets scientists adapt the model to jobs like running chemistry experiments, designing proteins and planning laboratory work, with CUDA-X tools handling the computation.

Early users show the range. Terray Therapeutics post-trained Nemotron to pull reactants, products and reaction conditions from scientific diagrams, recovering about 76% of annotated reactions on the RxnScribe test. Lila Sciences used it to design execution-ready laboratory protocols and nearly doubled its accuracy for making quantum dots, to 61%, ahead of GPT-6 Astra's 54% at 10 to 20 times lower inference cost. Therna nearly doubled accuracy in predicting protein-expression ranges for mRNA therapeutic design, with scientists preferring its explanations in 96% of comparisons.

Argonne National Laboratory is building Nemotron into its StructBioReasoner multi-agent system to design protein-based therapeutic candidates, a KTH Royal Institute of Technology agent fixed 81% of broken GROMACS molecular dynamics setups in a single overnight run on a DGX Spark, and RIKEN offers the models on its RIKYU supercomputing platform. The Nemotron 3.5 Super VL model and LitmusBench v0.2 are available now on Hugging Face, and the model is optimised to run locally on a 128GB DGX Spark system.