NVIDIA launched Nemotron for Science at SC26 on November 16 to help research teams adapt open AI models to chemistry, biology and laboratory workflows. Rather than adding another general-purpose foundation checkpoint, the initiative brings together the previously released Nemotron 3.5 Super VL, new scientific evaluation data and post-training techniques. The stated goal is agents that can interpret research materials, operate scientific tools and support controlled experiments on infrastructure a laboratory manages. Primary source ↗
The most concrete new release is LitmusBench v0.2, a chemistry resource with nearly 50,000 questions. NVIDIA says the material draws on molecular-property prediction and drug-discovery data, including tasks related to solubility and biological processing. The dataset is visible on Hugging Face. Its metadata contains dataset-specific licensing information, which should not be conflated with the downloadable model's OpenMDW 1.1 terms. Verifying data provenance and a pinned dataset revision remains important before reuse.
Nemotron 3.5 Super VL is a sparse vision-language model combining Mamba-2, attention and mixture-of-experts components with a vision encoder. The published BF16 card specifies 120 billion total parameters, approximately 12 billion activated parameters per token and up to a million-token context window. NVIDIA says supported versions can operate on DGX Spark-class systems as well as larger servers; performance and memory limits vary with quantization, context length and runtime rather than following from the total parameter count alone.
The company highlights partners adapting Nemotron to reaction extraction, laboratory-protocol development, protein expression and automatic repair of molecular-dynamics setups. These reports make the scientific direction credible, but their accuracy, latency and cost figures are vendor- or partner-reported under specialized experimental conditions. Open Model Weights does not claim to have repeated those tests. The development matters because scientific AI adoption requires more than a checkpoint: usable data, reproducible recipes and explicit model-versus-dataset terms shape what research teams can actually deploy.
What this changes for the evidence layer
Do not list Nemotron for Science as a new independent model family solely because a collection and recipes were announced. Track existing Nemotron 3.5 Super VL model card and its OpenMDW 1.1 download terms separately from LitmusBench v0.2 dataset metadata, revisions and conditions. Partner task-level improvements are publisher-reported, not independently measured.
NVIDIA Newsroom — Nemotron for Science at SC26
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗