The Pulse
Nvidia releases chemistry dataset with Nemotron for Science
Nvidia's Nemotron for Science combines an open model with a chemistry dataset and post-training recipes for scientific agents. The company says researchers are using the tools for tasks ranging from reaction extraction to therapeutic design.

Nvidia has released a chemistry dataset with nearly 50,000 questions as part of Nemotron for Science, a collection of tools for building specialized scientific AI agents. The package brings together the open Nemotron 3.5 Super VL model, scientific evaluations and post-training recipes, the company said in its November 16 announcement.
Source nvidianews.nvidia.com
LitmusBench v0.2 is intended for both training and evaluating language models on chemistry reasoning. The dataset and model are available on Hugging Face.
Adapting models to research tasks
Nvidia says researchers can post-train its open models with domain-specific data, then use them in agents that work through scientific problems and use research tools. The company lists chemistry experiments and protein design among the tasks for which scientists can adapt the model.
LitmusBench draws on molecular properties and experimental drug discovery data. Its questions concern predictions of molecular behavior, including solubility and how the human body processes molecules.
Researchers can integrate a post-trained model into workflows that interpret scientific information, generate code and call computational tools. Those workflows can use Nvidia CUDA-X accelerated tools for scientific computation.
Life sciences applications
Terray Therapeutics post-trained Nemotron to extract reactants, products and reaction conditions from scientific diagrams. Nvidia says it surpassed previously published results on two benchmarks, recovering about 76% of annotated reactions on the RxnScribe test under that benchmark's matching criteria.
Lila Sciences used Nemotron to design execution-ready laboratory protocols for its automated labs. Nvidia says post-training with laboratory feedback nearly doubled the model's accuracy for making quantum dots to 61%, compared with 54% for GPT-6 Astra, at 10-20x lower inference cost.
Quantum dots are nanoscale crystals used in screens and solar panels. NVIDIA says post-training with laboratory feedback nearly doubled the model’s accuracy for making quantum dots, and that such gains help agents and materials scientists translate desired material properties into synthesis recipes.
Therna worked with Nvidia to post-train Nemotron for agentic mRNA therapeutic design. According to Nvidia, the resulting model nearly doubled accuracy in predicting protein-expression ranges over frontier models, while scientists preferred its explanations to those of the base model in 96% of comparisons.
Research centers and access
Argonne National Laboratory is integrating Nemotron into StructBioReasoner, its multi-agent system for designing and evaluating protein-based therapeutic candidates for cancer and other diseases. The Jülich Supercomputing Centre is also integrating the model into its drug and scientific discovery AI agents.
KTH Royal Institute of Technology has built a Nemotron-based agent to repair broken GROMACS molecular dynamics setups. In one overnight run on a single Nvidia DGX Spark system, the agent fixed 81% of test cases with reasoning enabled, and GROMACS independently checked every fix, Nvidia says.
RIKEN has made Nemotron models, including Nemotron 3.5 Super VL, available on its RIKYU AI supercomputing platform through the RiVault inference gateway. Nemotron 3.5 Super VL was also optimized to run locally on a 128GB Nvidia DGX Spark system.
Source
Companies in this story
Related stories
- Nvidia claims Vera Rubin inference lead in provisional MLPerf results
- Nvidia RTX Spark desktops go on sale with up to 128GB memory
- Nvidia Reaffirms Korea Chip Plan After CEO Exit, TechCrunch Reports
- Nvidia’s private-company stakes reach $65.5 billion, TechCrunch reports
- Dyania and Nvidia team up on clinical AI inside hospital networks