NVIDIA releases open AI model for autonomous scientific research
The new Nemotron for Science platform allows labs to run complex chemistry and biology simulations on local hardware while keeping data private.

The short version
- Features the Nemotron 3.5 Super VL model with a 120 billion parameter Mixture-of-Experts architecture.
- Includes LitmusBench v0.2, a dataset of nearly 50,000 chemistry reasoning questions for model evaluation.
- Designed to run on local hardware, ranging from desktop workstations to large-scale supercomputing centers.
- Supports multimodal inputs like high-resolution figures, chemical reaction schemes, and instrumentation video.
Why it matters
This shift toward open-weights models allows researchers to conduct sensitive experiments without sending proprietary data to third-party cloud APIs. It also helps universities move from variable per-token costs to predictable capital investments in local hardware.
A new toolkit for labs
NVIDIA introduced Nemotron for Science during the SC26 supercomputing conference in Chicago. This initiative provides an open-weights AI model and specific training recipes to help scientists automate experimental design and molecular property prediction. It is specifically built for fields like chemistry, materials science, and molecular biology.
The release includes the Nemotron 3.5 Super VL foundation model alongside a new evaluation benchmark. This package allows university laboratories and research institutions to build specialized agents that can reason through complex scientific problems. These agents can synthesize literature and draft synthesis protocols autonomously.
The goal is to move beyond static structural predictions, which are common in current AI tools. Instead, the system focuses on active, iterative research loops that can refine hypotheses over time. As Ian Buck of NVIDIA stated, the tool gives researchers the means to turn ideas into discoveries using accelerated tools.
How the model works
The Nemotron 3.5 Super VL model uses a Mixture-of-Experts architecture. This means it has 120 billion total parameters, but only activates 12 billion parameters per token to process information. This design allows it to handle text, high-resolution figures, and even instrumentation video within a single pipeline.
By processing these different types of data together, the model avoids the latency and context loss that occurs when chaining separate vision and text models. It also uses several optimization techniques, such as LiLiCorr speculative decoding and NVFP4 quantization. These methods help it deliver high throughput, reaching over 600 tokens per second on a single GPU.
The system also includes LitmusBench v0.2, which contains nearly 50,000 chemistry reasoning questions. These questions are drawn from drug discovery records and are used to evaluate how well a model understands things like aqueous solubility and pharmacokinetic profiles. The benchmark is structured for Reinforcement Learning with Verifiable Rewards, which uses experimental outcomes as programmatic reward functions.
Impact on research privacy
Many academic and clinical research offices face strict constraints when using external commercial APIs. Transmitting novel chemical structures or pre-patent compounds to third-party endpoints creates significant intellectual property and confidentiality concerns. Because Nemotron for Science is an open-weights model, it can run entirely on local institutional clusters.
This on-premises approach ensures that researchers can maintain compliance with export controls and internal privacy policies. It also helps institutions manage their budgets more effectively by avoiding variable per-token cloud fees. Instead, research directors can shift grant expenditures toward predictable investments in local hardware like the NVIDIA DGX Spark workstation.
Several institutions have already begun testing these capabilities in various ways. For example, researchers at KTH Royal Institute of Technology built an agent to repair broken molecular dynamics simulation setups. This agent successfully resolved 81% of broken test cases while running overnight on a single local workstation.
Real-world applications
The platform is already being used to solve specific problems in the life sciences and materials science. Argonne National Laboratory is using the model to design protein therapeutic candidates for cancer. Meanwhile, RIKEN in Japan deployed the models across its supercomputing platform to provide secure, domestic access for researchers.
Private companies are also finding practical uses for the technology in their workflows. Lila Sciences used the model to generate laboratory protocols for quantum-dot synthesis, achieving 61% accuracy. Another company, Therna, used it for mRNA therapeutic design and reported a doubling in prediction accuracy for protein expression ranges.
Terray Therapeutics also utilized the model to extract chemical reaction conditions from literature diagrams. They were able to recover 76% of annotated reactions to build a structured synthesis database. These examples demonstrate how the model can be fine-tuned for specific tasks like extracting data from complex diagrams or generating automated protocols.
The path forward
NVIDIA has made the model weights and the LitmusBench dataset available on Hugging Face and NVIDIA NGC. The company plans to expand its NeMo Gym environments to include other scientific disciplines, such as genomics and crystallography. This will provide a broader foundation for different types of automated scientific research.
Independent academic consortia are currently preparing community benchmark evaluations across standard cheminformatics tasks. These evaluations will help establish verifiable methodologies for reproducible AI experiments across various university laboratories. This move supports the goal of making scientific AI more transparent and standardized for the public.
The release marks a significant shift toward agentic systems that can orchestrate simulations and refine hypotheses. By providing the tools for supervised fine-tuning and low-rank adaptation, NVIDIA is enabling researchers to create highly specialized tools. These tools will likely become standard for labs that require high-performance computing without the need for external cloud data transmission.
What to watch
- Expansion of NeMo Gym environments into genomics and crystallography.
- Results from independent academic consortia on community benchmark evaluations.
- More use cases for autonomous simulation debugging in molecular dynamics.
Try it
Check Hugging Face for the 'nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16' model weights to see the technical specifications.
Sources and links
- Officialnvidianews.nvidia.com/news/nvidia-introduces-nemotron-for-science-to-a nvidianews.nvidia.com
- Officialdeveloper.nvidia.com/blog/post-training-nvidia-nemotron-to-increase-ac developer.nvidia.com
- Officialdeveloper.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-speciali developer.nvidia.com
- Modelhuggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 huggingface.co
- ReportingCASRAI: the article we sourced
First reported by CASRAI. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.