At the SC26 supercomputing conference on November 16, 2026, NVIDIA announced NVIDIA Nemotron for Science, an AI development platform for scientific research. It is not a finished "AI scientist." Instead, it combines an open vision-language model, a chemistry dataset of about 50,000 questions and a reinforcement-learning post-training recipe, so that researchers can build AI agents specialized for their own fields. The announcement also highlighted results such as 61% accuracy in predicting quantum dot synthesis and an 81% repair rate for molecular dynamics configurations. However, the materials backing these figures vary, from technical blog posts to preprints to corporate press releases, and the extent to which each can be verified differs. Drawing on the published technical documents, this article sorts out what NVIDIA is providing and how far scientific AI agents have actually come.

AD

Not a chatbot, but a kit that researchers train themselves

Nemotron for Science is not a service for asking an AI science questions in conversation. It is designed so that researchers and companies can further train a base open model on their own specialist data and build it into AI agents that operate scientific computing tools. Ian Buck, NVIDIA's vice president of hyperscale and HPC, described it as a way to help researchers build specialized agents that reason about scientific problems and advance research using CUDA-X-accelerated tools, by providing open models, scientific data and post-training recipes.

The base model, NVIDIA Nemotron 3.5 Super-VL, is a vision-language model with 120B (120 billion) total parameters and 12B (12 billion) active parameters at inference. It uses a hybrid Mixture of Experts (MoE) architecture combining Mamba-2 and Transformer, and integrates the C-RADIO vision encoder for image processing. It accepts text, images and video as input and supports a context length of up to 1 million tokens.

It is released under the OpenMDW-1.1 license, which permits both commercial and non-commercial use. For running the standard BF16-precision model, NVIDIA lists a reference configuration of four H100 GPUs or two GB200 GPUs. NVIDIA is also working on optimizations for local execution on DGX Spark, a desktop AI computer with 128GB of memory. Note that Nemotron 3.5 Super-VL itself was not first released this time; it was published on Hugging Face on October 15, 2026, ahead of the SC26 announcement.

What's inside LitmusBench v0.2, the roughly 50,000-question chemistry dataset

As training data for scientific specialization, NVIDIA offers an open chemistry dataset called LitmusBench v0.2. The announcement describes it as about 50,000 questions; according to the dataset card, it breaks down into 47,118 training questions and 2,056 test questions, for a total of 49,174. The data comes from the public database ChEMBL (v36) plus datasets provided by ExpansionRx and Biogen, and some of it uses descriptors calculated with the chemistry library RDKit. It covers 126 physical and chemical properties, such as molecular solubility and metabolism, and 12 assay (biological evaluation) endpoints.

All questions are text-based short-answer items. For example, a SMILES string representing a molecular structure is given as input, and the model is asked to return a specific numerical value, such as a QED score that evaluates drug-likeness. The dataset card was created on September 2, 2026. ChEMBL-derived questions are licensed under CC BY-SA 4.0, and the rest under OpenMDW 1.1.

The predecessor, v0.1, was created on March 15, 2026 and consisted of 5,232 training questions and 482 test questions based on ChEMBL and RDKit descriptors. In v0.2, the training data grew roughly ninefold and real assay data was added. Its main use is "Reinforcement Learning with Verifiable Rewards" (RLVR) with NeMo Gym, in combination with training frameworks such as NeMo RL, TRL, Unsloth and veRL. Because it works by checking answers against recorded values for known molecules, it does not directly verify unknown experimental results.

AD

Using lab records for automatic grading in reinforcement learning

A key element of Nemotron for Science is post-training with RLVR on scientific data. Unlike RLHF, in which humans judge how good a response is, RLVR automatically grades a model's answers against predefined criteria and uses the result as a reward for learning. In a technical blog post published on November 16, NVIDIA explains concretely how to apply it to scientific fields.

In this approach, CSV files that record SMILES strings and measured values are converted into pairs of questions and correct answers for the model. The model's generated answers are compared with the recorded ground truth, and the degree of agreement is calculated as a reward, so reinforcement learning can proceed without humans evaluating each answer. The distinguishing feature is that existing experimental records can be turned directly into an automatic grading mechanism for training the model.

The blog post presents drug metabolism prediction using the metabolism database MetxBioDB as a concrete example. The model is given a parent compound and information about the enzyme involved in metabolism, and is asked to predict the structure of the resulting metabolite. The prediction is automatically graded against the correct molecular structure using Tanimoto similarity, which expresses molecular fingerprint similarity on a scale of 0 to 1.

Training uses GRPO (Group Relative Policy Optimization) with LoRA adapters, with NeMo Gym handling the training environment and NeMo RL handling the reinforcement learning. In this metabolism prediction task, with the output limit set to 16,384 tokens, accuracy reportedly improved step by step through supervised fine-tuning (SFT) followed by reinforcement learning. To prepare the initial SFT training data, NVIDIA also used a distillation approach in which Claude Sonnet 5.5 generated reasoning traces. From 1,934 candidates, it selected 1,123 whose Tanimoto similarity to the correct answer was 0.7 or higher and used them as training data.

This improvement is a result for a specific metabolite prediction task published on NVIDIA's technical blog, not a finding reported in a peer-reviewed paper. Also, the Tanimoto similarity used for grading measures structural similarity between molecules; it does not directly assess whether a predicted metabolite would actually be produced in a real experiment.

Note also that the technical blog calls the dataset "Litmus-Bench V2.0," while the official Hugging Face dataset card and the press release use "LitmusBench v0.2."

Results reported by partners, and how far they can be verified

In this announcement, NVIDIA also presented examples of Nemotron use by life sciences companies and research institutions. However, the published technical materials and evaluation methods differ for each result, so the evaluation target and metric must be distinguished when comparing the figures.

Entity / subject Reported figure Task and evaluation metric Source
NVIDIA (training recipe) Accuracy improved step by step through SFT and RL Drug metabolism prediction with MetxBioDB (Tanimoto similarity) NVIDIA technical blog
Terray Therapeutics HybridMatch F1 of 63.4 (Nemotron Omni). Super-VL also surpassed the prior model in initial evaluation Parsing chemical reaction diagrams with RxnCaption-15k and RxnScribe Terray official technical article
Terray Therapeutics Extracted about 76% of annotated reactions RxnScribe test (custom matching criteria) NVIDIA press release
Lila Sciences 61% accuracy (GPT-6 Astra: 54%) Designing quantum dot synthesis protocols NVIDIA press release
Therna Prediction accuracy nearly doubled; explanation preference rate of 96% Predicting protein expression range from mRNA Therna/NVIDIA announcement
Argonne National Laboratory 50.98% better than baseline by one standard deviation or more (Der f 21) Computational design of protein binders (MM-PBSA) Preprint arXiv:2512.15930
KTH Royal Institute of Technology Automatically repaired 81% of test cases Repairing broken GROMACS configurations NVIDIA press release

Terray Therapeutics has published relatively detailed validation data. In a technical article dated October 15, 2026, the company reports the performance of a model that extracts reactants, products and reaction conditions from chemical reaction diagrams. Against a HybridMatch F1 score of 59.8 for the prior work RxnCaption-VL, fine-tuning Nemotron Omni with bounding boxes from MolYOLO object detection and 5,000 synthetic reaction diagrams of its own creation achieved 63.4 on the RxnCaption-15k test set. The test set contains 400 diagrams and about 2,800 chemical reactions.

The company also reports that in an initial single-run evaluation of Nemotron 3.5 Super-VL, it surpassed RxnCaption-VL under conditions using RxnScribe bounding boxes. Meanwhile, the figure in NVIDIA's press release, that about 76% of annotated reactions were recovered in the RxnScribe test, is based on different evaluation criteria from the HybridMatch F1 that Terray reported in its technical article.

Regarding Lila Sciences, according to NVIDIA's press release, post-training the model with feedback from the lab nearly doubled the accuracy of quantum dot synthesis protocol design, reaching 61%. The comparison model, GPT-6 Astra, scored 54%, and inference cost was reportedly one-tenth to one-twentieth of that model's. However, neither the pre-training accuracy nor the absolute inference costs used in the comparison appear in the announcement materials we could confirm.

Therna, together with NVIDIA, developed BioReason-RNA, a model specialized for designing mRNA therapeutics. According to the company, its accuracy in predicting protein expression range was nearly double that of the previous state-of-the-art model, and scientists chose BioReason-RNA's answers in 96% of comparisons of the generated explanations. However, the models used for comparison, the absolute prediction accuracy and details of the evaluation data have not been disclosed. For both Lila Sciences and Therna, we could not find a published paper or technical report describing the specific test conditions or benchmark design.

AD

Research institutions' implementations, and the distance to experimental validation

Research institutions are also putting the technology to use, but most of the results reported so far rest on computational validation.

Argonne National Laboratory, part of the U.S. Department of Energy, has integrated Nemotron into StructBioReasoner, a multi-agent system that supports the design of proteins as candidate therapeutics for diseases such as cancer. The preprint (arXiv:2512.15930) underlying this work was first submitted on December 17, 2025, and a second version was released on April 27, 2026, making it research that predates NVIDIA's announcement.

The results reported in the paper rest on in silico (computer-based) evaluations, such as molecular dynamics simulations and binding free energy estimates using the MM-PBSA method. For example, in binder design for the allergen protein Der f 21, 50.98% of the 787 designs that passed quality control showed values at least one standard deviation below the reference binder's binding free energy of -145.25 kcal/mol. For the cancer-related protein NMNAT-2, 97,066 of the 266,606 candidates generated passed sequence and structure validation, and three main binding modes, including those involving the p53 interaction interface, were identified. The authors position these computational results as promising hypotheses and state explicitly that confirming the candidate proteins' properties and efficacy requires evaluation in the actual wet lab.

Sweden's KTH Royal Institute of Technology built an AI agent that automatically repairs broken configurations of the molecular dynamics software GROMACS by running a single DGX Spark overnight. According to NVIDIA's announcement, the agent repaired 81% of test cases and verified its fixes using GROMACS itself. This is an example using actual scientific computing software, but no peer-reviewed paper detailing the evaluation conditions or independent reproduction report has been published.

In Japan, RIKEN has begun offering Nemotron 3.5 Super-VL through RiVault, the inference gateway of its AI supercomputing platform RIKYU.

What distinguishes Nemotron for Science is that it goes beyond providing a science-oriented model: it opens up a mechanism that lets research institutions use their own experimental data for training and develop specialized agents independently. By combining model weights, a chemistry dataset of about 50,000 questions and a reinforcement-learning recipe built on automatic grading, researchers can tune the model to their own fields. Going forward, what will matter is how well the molecular structures and experimental protocols proposed by such agents can be reproduced in actual laboratories, not just evaluated on computers.