THE ESSENTIALS IN 30 SECONDS

  • The package combines the Nemotron 3.5 Super VL open model, the LitmusBench v0.2 chemistry dataset of nearly 50,000 questions, and recipes for post-training science agents.
  • Model weights and the dataset are on Hugging Face; the model is optimized to run locally on a 128GB DGX Spark.
  • Terray, Lila Sciences, Therna, Argonne, Jülich, KTH and RIKEN are named as users; their results are reported by Nvidia, not independently verified.

Nvidia used the SC26 supercomputing conference in Chicago on November 16 to present Nemotron for Science, a bundle meant to help researchers build AI agents specialized in scientific work. It is less a new model than a toolkit around an existing one.

What is in the package

The base is Nemotron 3.5 Super VL, an open multimodal model. According to Unite.AI, citing the model card, it was first published on Hugging Face in mid-October, has about 120 billion parameters with roughly 12 billion active, and accepts text, images and video. Nvidia says it is optimized to run locally on a DGX Spark with 128GB of memory.

The new pieces are LitmusBench v0.2, an open dataset of nearly 50,000 chemistry questions drawn from molecular property and experimental drug-discovery data (such as solubility and metabolism), and post-training recipes for adapting the model to tasks like chemistry experiments and protein design. Agents built this way can call Nvidia's CUDA-X accelerated scientific tools. The weights and the dataset are available on Hugging Face now. Nvidia's release does not state the licenses; Unite.AI lists CC-BY-SA-4.0 and OpenMDW-1.1 for the dataset.

Who is using it

Nvidia names several early users. Terray Therapeutics trained the model to pull reactions out of chemical diagrams and reports recovering about 76% of annotated reactions on the RxnScribe test. Lila Sciences says that after post-training on lab feedback, accuracy on quantum dot synthesis protocols rose to 61%, against 54% for a frontier model it identifies as GPT-6 Astra, at 10 to 20 times lower inference cost. Therna applied it to mRNA therapeutic design. KTH Royal Institute of Technology built an agent that repaired 81% of broken GROMACS molecular dynamics setups in one overnight run on a single DGX Spark.

On the public research side, Argonne National Laboratory is adding Nemotron to its StructBioReasoner multi-agent system for protein-based therapeutic candidates, Germany's Jülich Supercomputing Centre is integrating it into discovery agents, and Japan's RIKEN offers Nemotron models on its RIKYU platform through the RiVault inference gateway.

Caveats

All partner figures come from Nvidia's announcement and cover narrow, well-defined tasks; none had been independently replicated at launch. As Unite.AI notes, gains of this kind depend on how well the automatic scorer captures the scientific result researchers care about, and they do not show that agents can run a lab on their own.