NVIDIA introduced its latest tool for scientific applications, the Nemotron for Science, at the SC26 conference in Chicago on November 16, 2026. The technology, developed with the assistance of KTH Royal Institute of Technology, reportedly repaired 81% of intentionally broken GROMACS molecular-dynamics configurations within a single overnight run on a DGX Spark workstation. This accomplishment was confirmed by GROMACS rather than using any internal assessment method.
The announcement included a package consisting of three key artefacts: the open Nemotron 3.5 Super VL checkpoint, LitmusBench v0.2 (a dataset containing approximately 50,000 chemistry-reasoning questions), and post-training recipes for integrating the model with CUDA-X accelerated scientific tools.
The Three Artefacts
The three components offer distinct features and licensing arrangements:
- The checkpoint: Nemotron 3.5 Super VL features 120 billion parameters, with 12 billion being active per token. This hybrid model incorporates Mamba-2 and Transformer architectures and supports up to 1 million tokens of context. While the repository was available as of October 12, the November announcement highlighted its commercial and non-commercial utilization under the OpenMDW-1.1 license.
- The dataset: LitmusBench v0.2 offers a substantial resource for evaluating chemistry reasoning and can also serve as post-training data. It specifically addresses solubility prediction and metabolic fate tasks, making it a credible tool for research and development in these areas.
- The recipes and tooling: These provide essential guidance for adapting the model to specific scientific workflows, allowing the user to transition from a basic chat model to one capable of running simulations autonomously.
The announcement featured results from seven collaborating organizations, though NVIDIA emphasized that these numbers should be viewed as case studies rather than definite metrics due to their vendor-reported nature.
Among the highlights, Terray Therapeutics post-trained the model to extract details from chemistry diagrams, reportedly achieving a 76% accuracy rate on annotated reactions. Meanwhile, Lila Sciences noted improved performance in generating automated lab protocols, reaching a quantum-dot synthesis accuracy of 61% when compared to prior benchmarks.
Notably, the integration of Nemotron into existing systems at institutions like Argonne National Laboratory and RIKEN was also part of the announcement, showcasing its practical application in protein-based drug development and other areas.
While the results and frameworks appear promising, there exist gaps in the release. There was no independent evaluation of the science claims, no published pricing for the model as a service, and no information regarding the operational efficiency of the agent loop that would clarify its viability for everyday research tasks.
Moving forward, attention will be directed toward forthcoming independent assessments of the LitmusBench v0.2 dataset, a potential pricing model for the service, and detailed reports on the efficiency of the agent loop. These insights are crucial for organizations considering the adoption of this tool for scientific workflows.


