NVIDIA launches Nemotron for Science to boost AI agents in research
NVIDIA unveiled Nemotron for Science, pairing its open Nemotron 3.5 Super VL model with a new chemistry dataset and post‑training recipes.
Source: NVIDIA Press RoomWhat is NVIDIA Nemotron for Science?
NVIDIA announced a new offering called Nemotron for Science, which bundles the open‑source Nemotron 3.5 Super VL model, a chemistry‑focused benchmark called LitmusBench v0.2, and a set of post‑training recipes. The model itself is an open‑weight large language model that can be run on NVIDIA DGX Spark systems or accessed via Hugging Face, according to the NVIDIA Press Room. The post‑training recipes let scientists fine‑tune the base model with domain‑specific data so that the resulting agent can reason about experiments, generate code, and call CUDA‑X‑accelerated scientific tools.
Which organizations are using Nemotron for Science and how?
Early adopters span both industry and national labs. Life‑science firms Lila Sciences, Terray Therapeutics, and Therna are customizing Nemotron 3.5 for tasks such as extracting reaction details from diagrams, designing execution‑ready lab protocols, and engineering mRNA therapeutic sequences. Research centers including the U.S. Department of Energy’s Argonne National Laboratory, Germany’s Jülich Supercomputing Centre, and RIKEN have integrated the model into multi‑agent systems that design protein therapeutics, repair molecular‑dynamics setups, and provide on‑demand inference through supercomputing gateways. These deployments illustrate the model’s flexibility across chemistry, protein design, and materials synthesis.
How does LitmusBench v0.2 support chemistry reasoning?
LitmusBench v0.2 is an open dataset released alongside Nemotron for Science. It contains nearly 50,000 question‑answer pairs that draw on molecular properties, solubility data, and experimental drug‑discovery metrics. Researchers can use the benchmark both to evaluate a model’s chemistry understanding and as a training source for post‑training. By measuring performance on tasks such as predicting solubility or metabolic pathways, scientists can gauge how well a fine‑tuned agent will translate raw literature into actionable experimental plans.
What performance improvements have been demonstrated?
The press release cites several concrete gains. Terray Therapeutics post‑trained Nemotron to extract reactants, products, and conditions from scientific diagrams, achieving about 76 % recovery on the RxnScribe benchmark—higher than previously published results. Lila Sciences reported that after a feedback loop with its automated labs, the model’s accuracy for designing quantum‑dot synthesis protocols rose to 61 %, surpassing GPT‑6 Astra’s 54 % while using 10‑20× lower inference cost. Therna claimed a near‑doubling of accuracy in predicting protein‑expression ranges for mRNA therapeutics, and scientists preferred the model’s explanations over the base version in 96 % of head‑to‑head comparisons. At the KTH Royal Institute of Technology, an agent built on Nemotron repaired 81 % of broken GROMACS molecular‑dynamics cases in an overnight run, with every fix validated by GROMACS itself.
How can creators, developers, and researchers experiment with Nemotron using Grafy?
Grafy’s Chat app lets users spin up multi‑model conversations, including models hosted on Hugging Face. By selecting the Nemotron 3.5 Super VL checkpoint, a researcher can interact with the base model directly in the browser, testing its chemistry reasoning on the fly. For deeper investigations, the Reader app can ingest the LitmusBench v0.2 PDF or CSV files, enabling keyword search, summarisation, and question‑answering over the dataset without leaving the workspace. Developers who want to try the post‑training recipes can use Grafy Code to run Python notebooks that pull the dataset, apply LoRA‑style fine‑tuning, and export a fine‑tuned checkpoint ready for deployment on a local DGX Spark or a cloud instance. Because all these apps share a common workspace, users can prototype an end‑to‑end scientific agent: query the model in Chat, retrieve relevant literature in Reader, generate synthesis scripts in Code, and even visualise molecular structures with the Creator image generator if desired.
What are the broader implications for AI‑driven scientific discovery?
Nemotron for Science illustrates a growing trend: open‑weight models coupled with domain‑specific datasets and easy‑to‑apply fine‑tuning recipes are lowering the barrier for labs to build their own AI assistants. By making the model and benchmark openly available, NVIDIA encourages reproducibility and community‑driven improvements, which could accelerate breakthroughs in drug design, materials science, and even climate modeling. The integration with CUDA‑X tools means that once an agent proposes a computational experiment, the same hardware that powers the model can execute the simulation, creating a tight feedback loop between reasoning and execution. As more institutions adopt this workflow, we may see a shift from isolated AI pilots to production‑grade scientific agents that operate across the entire research lifecycle.
Frequently asked questions
How does post‑training differ from fine‑tuning for Nemotron?
Post‑training in NVIDIA’s context refers to applying lightweight adaptation methods—such as LoRA or adapter layers—on top of the frozen base model using domain‑specific data. This approach preserves the original weights, reduces compute cost, and allows the same checkpoint to serve multiple specialized agents, unlike full fine‑tuning which rewrites the entire model.
Can I run Nemotron 3.5 locally without a DGX system?
Yes. NVIDIA optimized the model to run on a single 128 GB DGX Spark system, but the checkpoint is also hosted on Hugging Face and can be executed on any GPU‑enabled machine that meets the memory requirements. Grafy Chat can connect to the hosted version for quick experiments.
Is the LitmusBench dataset suitable for training models beyond chemistry?
LitmusBench v0.2 focuses on chemistry reasoning, but its structure—question‑answer pairs linked to real‑world molecular data—makes it a useful template for other scientific domains. Researchers can adapt the data format to create similar benchmarks for materials science, biology, or physics, leveraging the same post‑training pipelines.
Source: NVIDIA Press Room — NVIDIA Introduces Nemotron for Science to Accelerate Discoveries With Agents


