Nvidia has launched Nemotron for Science, an open package for building AI agents that handle chemistry and lab work. The early results come from Nvidia and its partners.

Nvidia launched Nemotron for Science on 16 November, a package that pairs its open Nemotron 3.5 Super VL model with a new chemistry dataset and recipes for building AI agents that can work through scientific problems

1

2

.

Source: NVIDIA

Source: NVIDIA

The release builds on the existing Nemotron 3.5 Super VL open model, adapted here for life sciences and chemistry

1

. Its open artifact is LitmusBench v0.2, a dataset of nearly 50,000 questions for training and testing language models on chemistry reasoning, drawn from molecular properties and experimental drug discovery data

1

2

. The model and the dataset are on Hugging Face now, and the model was optimized to run locally on a 128GB DGX Spark system

1

2

. Nvidia says accelerated computing is changing how scientists tackle hard problems, and that open models, data and recipes let them turn ideas into discoveries

1

. "Nemotron for Science gives researchers open models, scientific data and post-training recipes to build specialized agents that reason through scientific problems and use CUDA-X accelerated tools to turn ideas into discoveries," said Ian Buck, the company's vice president of hyperscale and high-performance computing

1

. The agents can read scientific information, write code and call computational tools

1

.

What the early results show

Three life sciences companies report results from post-training the model. The partners are using it to pull chemical knowledge out of data, plan experiments and design therapeutics

1

. Terray Therapeutics post-trained the model to pull reactants, products and reaction conditions out of scientific diagrams. On the RxnScribe test, the company says, it recovered about 76 percent of the annotated reactions, judged by the benchmark's matching criteria. By the company's own account, that tops previously published results on the two benchmarks

1

.

Lila Sciences used the model to write execution-ready protocols for its automated labs, and says it nearly matches frontier performance without extra fine-tuning

1

. On one task, synthesizing quantum dots, post-training with feedback from its labs nearly doubled accuracy, to 61 percent against GPT-6 Astra's 54 percent. The release claims that came at 10 to 20 times lower inference cost

1

. Therna worked with Nvidia to post-train the model for designing mRNA therapeutics. It says the model nearly doubled its accuracy at predicting protein-expression ranges over frontier models.

From desktops to supercomputers

Research institutions are picking the model up too. Argonne National Laboratory, a U.S. Department of Energy lab, is integrating Nemotron into its StructBioReasoner system, which designs and evaluates protein-based therapeutic candidates for cancer and other illnesses

1

. KTH Royal Institute of Technology built an agent that repairs broken GROMACS molecular dynamics setups. In one overnight run on a single DGX Spark it fixed 81 percent of test cases, each one checked by GROMACS itself

1

.

Supercomputing centers are in the picture too. Riken has put Nemotron on its Rikyu AI supercomputing platform, served through its RiVault inference gateway. Germany's Jülich Supercomputing Centre is integrating it into its own discovery agents

1

. Nvidia positions the model as spanning scales from a researcher's desktop to a supercomputing center, running on infrastructure the organizations themselves control

1

. The company says the work could open new paths in critical materials, national security and cancer treatment

1

.

Every performance figure in the release comes from Nvidia or its partners, and no independent evaluation of them appears in the coverage

1

. Until an outside lab publishes its own numbers, the evidence for those results begins and ends with Nvidia's announcement

1

.

© 2026 TheOutpost.AI All rights reserved