NVIDIA's Nemotron for Science pairs open model with chemistry bench
NVIDIA's Nemotron for Science ships with LitmusBench v0.2, nearly 50,000 chemistry questions, and recipes that use experimental outcomes as automated rewards instead of human preference labels.

Why it matters
- NVIDIA launched Nemotron for Science on November 16, 2026, with LitmusBench v0.2, a chemistry dataset of nearly 50,000 questions.
- The underlying Nemotron 3.5 Super VL model has 120 billion total parameters, 12 billion active, and a one-million-token context window, released October 15, 2026.
- NVIDIA reports supervised fine-tuning improved drug-metabolism accuracy by 12%, with reinforcement learning adding a further 17%.
- Lila Sciences reports 61% accuracy on quantum-dot protocol generation versus GPT-6 Astra's 54%, at 10–20 times lower inference cost.
- A KTH-built agent fixed 81% of broken GROMACS molecular-dynamics setups in one overnight run on a single DGX Spark.
NVIDIA launched Nemotron for Science on November 16, 2026, pairing its open Nemotron 3.5 Super VL model with scientific evaluations, post-training recipes, and LitmusBench v0.2 — a chemistry dataset of nearly 50,000 questions for training and evaluating research agents. The package targets agents that interpret scientific information, generate code, and call computational tools, including CUDA-X accelerated software. Researchers can post-train the model with domain-specific data for work such as chemistry experiments and protein design, adapting a general vision-language system to a specific scientific task.
The underlying model is not the news here. Its model card dates the Hugging Face release to October 15, 2026, and describes 120 billion total parameters with 12 billion active, text, image, and video inputs, text outputs, and a context window of up to one million tokens. The new substance is the scientific data, the evaluations, and the customization recipes layered on top. The model and dataset are available on Hugging Face, and NVIDIA has optimized the model to run locally on a 128GB DGX Spark system.
Why does verification anchor the whole design?
LitmusBench v0.2 draws on molecular properties and experimental drug-discovery data, covering behavior such as solubility and how the human body processes molecules. The dataset repository identifies English text-generation data tagged for chemistry, molecular-property prediction, and reinforcement learning with verifiable rewards. It lists CC-BY-SA-4.0 and OpenMDW-1.1 licenses.
An NVIDIA technical guide explains the recipe, and the mechanism is straightforward. Experimental CSV records become question-answer pairs, with their measured outcomes serving as an automated verifier. That creates a reinforcement-learning environment without human preference labeling. Cheminformatics tools, molecular simulators, or a live autonomous laboratory can supply the scoring layer.
The guide's drug-metabolism example makes this concrete. The model receives a parent molecule and an enzyme, then predicts the resulting metabolite. A zero-to-one Tanimoto score measures overlap between the predicted and reference molecular fingerprints, producing a reward based on structural similarity. NVIDIA reports that supervised fine-tuning improved accuracy by 12% in this example, and reinforcement learning added a further 17% at a 16,384-token output budget.
There's an optional warm-start too. A stronger teacher generates reasoning traces, and the system keeps only those that meet the similarity threshold before reinforcement learning begins. The example in the guide uses Anthropic's Claude Sonnet 5.5 as that teacher. In short: the dataset turns measured experimental outcomes into automated rewards, which removes the need for humans to label preferences.
What do the life-sciences results actually show?
The announced applications span structured chemical extraction, laboratory-protocol generation, and therapeutic design. One caveat matters up front: these NVIDIA-reported results concern bounded tasks and do not establish general laboratory autonomy.
Terray Therapeutics post-trained Nemotron to extract reactants, products, and reaction conditions from scientific diagrams. NVIDIA reports that it recovered about 76% of annotated reactions on the RxnScribe test under that benchmark's matching criteria. That helps turn published chemistry into structured data for synthesis planning and scientific agents.
Lila Sciences used Nemotron to design execution-ready protocols for its automated laboratories. According to NVIDIA, post-training with laboratory feedback nearly doubled accuracy for making quantum dots — nanoscale crystals used in screens and solar panels — to 61%, compared with GPT-6 Astra's 54%, at 10–20 times lower inference cost.
Therna collaborated with NVIDIA on agentic mRNA therapeutic design. According to NVIDIA, the tuned model achieved nearly twice the accuracy of frontier models in predicting protein-expression ranges, and scientists preferred its explanations to the base model's in 96% of comparisons. The numbers come from the vendor, and each covers a narrow, well-scored task.
Who is deploying it, and on what hardware?
The research deployments sketch the practical picture:
- Argonne National Laboratory is integrating Nemotron into StructBioReasoner, a multi-agent system for designing and evaluating protein-based therapeutic candidates for cancer and other diseases.
- Jülich Supercomputing Centre is integrating the model into scientific-discovery agents.
- KTH Royal Institute of Technology built a Nemotron agent to repair broken GROMACS molecular-dynamics setups. NVIDIA reports that it fixed 81% of test cases with reasoning enabled during one overnight run on a single DGX Spark, with every fix independently checked by GROMACS.
- RIKEN has made Nemotron models, including 3.5 Super VL, available on its RIKYU AI supercomputing platform through the RiVault inference gateway for scientific research and accelerated-application development.
Alongside local DGX Spark deployments, the RIKEN route gives research organizations a way to run scientific workflows on infrastructure they control. For labs handling proprietary or sensitive data, running a 120-billion-parameter model on a local box rather than a commercial API is a meaningful option.
What are the limits of benchmark gains?
The unifying design across all of this is verification. The open model supplies reasoning and tool use, while experimental references, laboratory feedback, or domain software score its outputs. That division of labor is what lets NVIDIA skip human preference labeling entirely.
But it also concentrates the risk. Task selection, the quality of the scoring method, and held-out evaluation become central to the package's usefulness. A benchmark gain matters only when the verifier measures the scientific outcome researchers actually want. A Tanimoto score rewards structural similarity, not clinical validity; a lab-feedback loop rewards what that specific assay measures.
The stakes are real for both AI vendors and research institutions. Scientific AI has struggled with outputs that look plausible but fail experimental checks, and verifiable-reward training is one of the few mechanisms that ties model behavior directly to measured outcomes. NVIDIA is betting that an open model plus open data plus an open recipe will pull research labs into its hardware and CUDA-X software ecosystem. How far the approach travels beyond chemistry's tidy CSV records will decide whether this is a template for scientific AI or a well-instrumented niche.
Source: unite.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- NVIDIA Ships Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
- Nvidia launches open-weight Nemotron models for enterprise agents and AI PCs
- Nvidia's Nemotron 3.5 Super VL posts 5x agentic speed lead over open rivals
- NVIDIA's DGX Spark HPC Kit Puts a Four-Node Cluster on the Desk
- NVIDIA releases Alpamayo 2 Edge weights for autonomous driving