NVIDIA launches Nemotron for Science with open chemistry dataset LitmusBench
NVIDIA releases Nemotron 3.5 Super VL model, LitmusBench v0.2 dataset with 50,000 chemistry questions, and post-training recipes for scientific agents.

By TechWhoKnows
, updated 4 min read
NVIDIA has introduced Nemotron for Science, a collection of open models, datasets and post-training recipes designed to help researchers build specialised AI agents for scientific discovery. The announcement, made on 16 November 2026, centres on the Nemotron 3.5 Super VL vision-language model and a new open chemistry dataset called LitmusBench v0.2.
What Nemotron for Science includes
The platform brings together three components:
Nemotron 3.5 Super VL, an open vision-language model optimised to run locally on a 128 GB NVIDIA DGX Spark system
LitmusBench v0.2, a chemistry reasoning dataset with nearly 50,000 questions drawn from molecular properties and experimental drug discovery data, covering solubility prediction, metabolic processing and reaction understanding
Post-training recipes, documented methods for adapting the base model to specialised scientific tasks such as chemistry, protein design and materials science
All components are available on Hugging Face starting 16 November 2026.
Early results from research partners
Three life-sciences companies reported measurable gains from post-training Nemotron 3.5 Super VL:
Terray Therapeutics used the model to extract reactants, products and reaction conditions from scientific diagrams, recovering about 76% of annotated reactions on the RxnScribe benchmark under its matching criteria, surpassing previously published results.
Lila Sciences applied post-training with laboratory feedback to design execution-ready protocols for automated labs. Accuracy for quantum dot synthesis nearly doubled to 61%, exceeding GPT-6 Astra's 54% at 10, 20× lower inference cost.
Therna collaborated with NVIDIA on agentic mRNA therapeutic design, nearly doubling accuracy in predicting protein-expression ranges over frontier models. Scientists preferred its explanations over the base model in 96% of comparisons.
Research institution adoption
Major research centres are integrating Nemotron into their workflows:
Argonne National Laboratory (US Department of Energy) is incorporating Nemotron into its StructBioReasoner multi-agent system for designing and evaluating protein-based therapeutic candidates for cancer and other diseases.
KTH Royal Institute of Technology built an AI agent on Nemotron that repairs broken GROMACS molecular dynamics setups autonomously. In one overnight run on a single DGX Spark, the agent fixed 81% of test cases with reasoning enabled, and every fix was independently verified by GROMACS.
RIKEN has made Nemotron 3.5 Super VL available on its RIKYU AI supercomputing platform through RiVault, its inference gateway.
Why it matters
Scientific discovery increasingly depends on AI agents that can reason through domain-specific problems, generate code and call computational tools. By releasing an open model, a purpose-built evaluation dataset and documented post-training recipes together, NVIDIA is lowering the barrier for research teams to build and evaluate specialised agents without starting from scratch. The LitmusBench dataset in particular addresses a gap: chemistry reasoning benchmarks have been scarce compared to general-language benchmarks, making it difficult to measure progress on scientific tasks.
The ability to run Nemotron 3.5 Super VL on a single DGX Spark (128 GB) also means individual researchers and smaller labs can experiment locally before scaling to supercomputing centres, a practical consideration for adoption beyond the best-resourced institutions.
Sources
- 1NVIDIA Newsroom press releasenvidianews.nvidia.com