Skip to main content
v2026.11,986 entries · CC-BY 4.0

Editorial · CASRAI · AI and ML research outputs

NVIDIA Unveils Nemotron for Science and LitmusBench v0.2 at SC26

At SC26, NVIDIA launched Nemotron for Science and LitmusBench v0.2, open AI tools and chemistry datasets for autonomous scientific research.

Published 16 Nov 2026· 6 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

At the SC26 supercomputing conference in Chicago on November 16, 2026, NVIDIA introduced Nemotron for Science, an open artificial intelligence initiative designed to power autonomous agents for chemistry, materials science, and molecular biology. The release combines the multimodal Nemotron 3.5 Super VL foundation model, domain-specific post-training recipes, and LitmusBench v0.2—an open benchmark of nearly 50,000 chemistry reasoning questions—giving university laboratories and research institutions open tools to automate experimental design, literature synthesis, and molecular property prediction on local infrastructure.

What happened

On November 16, 2026, NVIDIA announced Nemotron for Science during the opening of the SC26 high-performance computing conference in Chicago. The announcement pairs an open-weights 120-billion-parameter multimodal model, Nemotron 3.5 Super VL, with domain-specific training environments and chemistry evaluation benchmarks released on Hugging Face. The package shifts computational focus from static structural predictions toward active, iterative multi-agent research loops.

“Accelerated computing is transforming how scientists tackle the world’s most complex challenges,” said Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA. “Nemotron for Science gives researchers open models, scientific data and post-training recipes to build specialized agents that reason through scientific problems and use CUDA-X accelerated tools to turn ideas into discoveries.”

NVIDIA designed the system to execute scientific workflows across diverse hardware tiers, scaling from a desktop 128GB NVIDIA DGX Spark workstation up to institutional clusters and leadership-class supercomputing centers without requiring external cloud data transmission.

Key details: models, datasets, and architecture

Nemotron for Science integrates three core components: an open multimodal vision-language model, an empirical chemistry benchmark, and structured reinforcement learning workflows.

Nemotron 3.5 Super VL: The model uses a Mixture-of-Experts (MoE) architecture with 120 billion total parameters, activating 12 billion parameters per token. It processes text, high-resolution scientific figures, chemical reaction schemes, and instrumentation video within a single pipeline, eliminating the context loss and latency of chaining separate text and computer vision models. Inference optimizations include LiLiCorr speculative decoding, NVFP4 quantization with distillation, multi-token prediction, and context reuse. On a single GPU, it delivers over 600 tokens per second on agentic benchmarks, offering up to 5x higher throughput than comparable open models.

LitmusBench v0.2: An open dataset of nearly 50,000 chemistry reasoning questions drawn from experimental drug discovery records and molecular property databases. It evaluates models on aqueous solubility, pharmacokinetic profiles, and absorption, distribution, metabolism, and excretion (ADME) pathways. Structured for Reinforcement Learning with Verifiable Rewards (RLVR), LitmusBench uses experimental outcomes as programmatic reward functions, removing the need for subjective human preference labeling.

Post-training recipes: Packaged via NVIDIA NeMo Gym and NeMo RL, the workflows guide supervised fine-tuning (SFT) and low-rank adaptation (LoRA) before reinforcement learning. In drug-metabolism benchmarks, NVIDIA reported a 12% accuracy gain from supervised fine-tuning and a further 17% gain from reinforcement learning under a 16,384-token budget. An optional distillation recipe uses frontier teacher models to seed initial reasoning traces before RL.

Background and industry context

The release marks an evolution beyond static structural prediction tools like AlphaFold and ESMFold toward agentic systems that orchestrate simulations, refine hypotheses, and draft synthesis protocols. While closed commercial models show strong general reasoning, academic and clinical research offices face strict constraints with external commercial APIs. Transmitting novel chemical structures or pre-patent compounds to third-party endpoints raises intellectual property and confidentiality concerns, while per-token API charges become prohibitive during iterative, overnight simulation campaigns.

By releasing open weights and code under permissive terms, NVIDIA allows universities to retain complete on-premises control over computational pipelines. This approach supports rigorous standards for computational reproducibility and establishes verifiable methodologies for every reproducible AI experiment conducted across university laboratories.

Adoption and real-world results

Several major research institutions and life sciences companies reported early results using the platform:

  • Argonne National Laboratory: Integrating Nemotron 3.5 Super VL into its StructBioReasoner multi-agent platform to design and evaluate protein therapeutic candidates for cancer and other complex diseases.
  • KTH Royal Institute of Technology: Researchers in Sweden built an autonomous agent to diagnose and repair broken GROMACS molecular dynamics simulation setups. Running overnight on a single 128GB DGX Spark workstation, the agent autonomously resolved 81% of broken test cases, with each configuration fix verified by GROMACS.
  • RIKEN: Japan’s national institute deployed Nemotron models across its RIKYU AI supercomputing platform via the RiVault inference gateway, enabling domestic researchers to access secure on-premises scientific agents.
  • Terray Therapeutics: Post-trained Nemotron to extract chemical reaction conditions and reactants from literature diagrams, recovering 76% of annotated reactions on the RxnScribe benchmark to build structured synthesis databases.
  • Lila Sciences: Used the model to generate automated laboratory protocols, achieving 61% accuracy in quantum-dot synthesis (outperforming GPT-6 Astra’s 54%) while reducing inference costs by 10x to 20x.
  • Therna: Post-trained Nemotron for mRNA therapeutic design, doubling prediction accuracy for protein expression ranges over baseline models, with research scientists preferring its mechanistic explanations in 96% of blind evaluations.
  • Jülich Supercomputing Centre: Deploying the stack within European research consortia to support automated simulation steering and materials exploration.

What it means for researchers, universities, and research offices

The emergence of capable open scientific foundation models carries tangible implications for academic research administrators:

1. Data governance and IP protection: Because model weights can run on local institutional clusters or departmental workstations, researchers working on patentable molecular scaffolds or sensitive clinical datasets can avoid third-party cloud APIs, maintaining compliance with export controls and intellectual property policies.

2. Predictable research computing budgets: Demonstrations from KTH and Lila Sciences show that domain-tuned 120B MoE models can run effectively on local workstations and campus clusters. This enables research computing directors to shift grant expenditures from variable per-token cloud operating fees into predictable capital equipment investments.

3. Open science and verifiable benchmarks: LitmusBench v0.2 provides an open standard for chemical reasoning evaluation. Research offices supporting compliance with public-access mandates can leverage these resources to align with FAIR data principles and contribute verified evaluation data to institutional open data repositories.

4. Federal grant sustainability: Investigators submitting proposals to the NSF, NIH, or Department of Energy can propose local, reproducible agentic pipelines without depending on external commercial API availability or terms of service.

What happens next

Model weights for Nemotron 3.5 Super VL and the LitmusBench v0.2 dataset are available immediately on Hugging Face and NVIDIA NGC. During SC26, NVIDIA and partners from Argonne and KTH are presenting live demonstrations of automated simulation debugging and chemical property prediction.

NVIDIA stated it will expand NeMo Gym environments to additional scientific disciplines, including crystallography and genomics, while independent academic consortia prepare community benchmark evaluations across standard cheminformatics tasks.

Sources

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →