NVIDIA has introduced Nemotron for Science, a collection of tools for developing AI agents focused on scientific research. It combines the open Nemotron 3.5 Super VL model, a new chemistry dataset and post-training recipes. Researchers can adapt the model for tasks such as chemistry experiments or protein design, according to the company.
A dataset for chemistry reasoning
The open LitmusBench v0.2 dataset contains nearly 50,000 questions for training and evaluating language models on chemistry reasoning. It draws on molecular properties and experimental drug-discovery data, including predictions of solubility and how the human body processes molecules.
According to NVIDIA, a model post-trained on domain-specific data can be incorporated into agent workflows that interpret scientific information, generate code and call computational tools. Those tools can include software accelerated with CUDA-X.
From chemistry diagrams to laboratory protocols
NVIDIA says Terray Therapeutics post-trained the model to extract reactants, products and reaction conditions from scientific diagrams. In the RxnScribe test, it recovered approximately 76% of annotated reactions under the benchmark's matching criteria.
Lila Sciences used Nemotron to design protocols ready to run in automated laboratories, according to NVIDIA. Without additional fine-tuning, it achieved performance comparable to frontier models.
NVIDIA says subsequent post-training with laboratory feedback nearly doubled the model's accuracy for making quantum dots to 61%. GPT-6 Astra reached 54%, while Nemotron's inference costs were 10–20 times lower.
Drug design and simulation repairs
Therna worked with NVIDIA to post-train Nemotron for agentic design of mRNA therapeutics. NVIDIA says the model nearly doubled accuracy in predicting protein-expression ranges compared with frontier models. Scientists preferred its explanations to those of the base model in 96% of comparisons.
Argonne National Laboratory is integrating Nemotron into StructBioReasoner, a multi-agent system designed to develop and evaluate protein-based drug candidates for cancer and other diseases.
KTH Royal Institute of Technology built a Nemotron-based agent to independently repair broken GROMACS molecular dynamics setups. According to NVIDIA, it fixed 81% of test cases with reasoning enabled during one overnight run on a single DGX Spark system. GROMACS itself checked every repair.
Model and dataset availability
Nemotron 3.5 Super VL and LitmusBench v0.2 are available on Hugging Face. The model is also optimized to run locally on an NVIDIA DGX Spark system with 128 GB of memory.
RIKEN has made Nemotron models, including 3.5 Super VL, available on its RIKYU supercomputing platform. Researchers can access them through the RiVault inference gateway for scientific research and the development of accelerated applications.



