NVIDIA announced the launch of Nemotron for Science — a toolkit built on the open Nemotron 3.5 Super VL model. The new offering combines the model, scientific benchmarks, an open chemistry dataset, and post-training recipes.
The goal is to help researchers build specialized AI agents for scientific tasks. The agents will be able to reason about scientific problems and use specialized research tools.
A key element is LitmusBench v0.2. It is an open dataset with nearly 50 thousand questions for training and evaluating models on chemical reasoning. The data are based on molecular properties and experimental data on drug development.
The model and tools run from the NVIDIA DGX Spark desktop system to supercomputing centers. This allows organizations to conduct research on their own infrastructure.
Several institutes and companies are already using the solution. Among them are Argonne National Laboratory, Jülich Supercomputing Centre, Lila Sciences, Terray Therapeutics, and Therna.
In one example, post-training with feedback from the laboratory raised the accuracy of quantum dot synthesis to 61% — higher than the result of GPT-6 Astra (54%) at 10–20 times lower inference costs.
Ian Buck, NVIDIA's vice president of hyperscale and high-performance computing, noted that accelerated computing is changing how scientists approach complex problems. Nemotron for Science gives researchers open models, data, and recipes for building agents that connect ideas with discoveries.
Only time will tell how quickly such agents will become part of laboratories' everyday work.

