NVIDIA Launches Nemotron for Science Tools
The open AI toolkit pairs a new chemistry dataset with post-trained models that life sciences firms are already customizing for research.
Updated on Nov. 16, 2026 in Artificial Intelligence

NVIDIA introduced Nemotron for Science, bringing together the NVIDIA Nemotron 3.5 Super VL open model, a new open chemistry dataset and post-training recipes. Life sciences companies Lila Sciences, Terray Therapeutics and Therna are using and customizing the model for scientific applications.
Why it matters
The release makes domain-specific scientific AI models available to researchers rather than only large proprietary labs, which could speed work in drug discovery and chemistry. Open availability also lets outside teams test and adapt the models for their own experiments.
LitmusBench v0.2 is an open dataset of nearly 50,000 chemistry reasoning questions drawn from molecular properties and experimental drug discovery data. On the RxnScribe test, a Terray Therapeutics model recovered about 76% of annotated reactions under the benchmark's matching criteria, and the source does not say how it performs on reactions outside that test.
The players
NVIDIA
NVIDIA is a U.S. technology company known for graphics processors and AI computing hardware and software, including the Nemotron family of open models.
Lila Sciences
Lila Sciences is a life sciences company that post-trained Nemotron with laboratory feedback to improve quantum dot manufacturing predictions.
Terray Therapeutics
Terray Therapeutics is a drug discovery company that post-trained Nemotron to read reactions from scientific diagrams.
Therna
Therna is a life sciences company that post-trained Nemotron for agentic mRNA therapeutic design.
The details
Researchers post-train the open Nemotron model with domain-specific data and integrate it into agent workflows that interpret scientific information, generate code and call computational tools. Terray Therapeutics used it to extract reactants, products and reaction conditions from scientific diagrams, while Therna post-trained it for agentic mRNA therapeutic design.
Timeline
November 16, 2026: NVIDIA introduces Nemotron for Science.
Under the Hood
The release extends NVIDIA's open Nemotron model family into scientific domains by pairing the base model with domain-specific post-training and datasets. It positions open, customizable models as an alternative to closed frontier systems, with Lila Sciences reporting higher accuracy than GPT-6 Astra on one task.
Researchers and developers can now download the Nemotron 3.5 Super VL model and the LitmusBench v0.2 dataset from Hugging Face to test or adapt them for their own work. Teams without access to proprietary frontier models can use these open tools to build domain-specific chemistry and biology applications.
The takeaway
Open, domain-tuned AI models are becoming practical tools for scientists, not only for the largest technology labs. Reviewing published benchmark methods before adopting these models in research will help teams judge how well results transfer to their own work.
Further reading
For more on the models driving these advances, see Artificial Intelligence coverage.
More information
Download NVIDIA Nemotron 3.5 Super VL and LitmusBench v0.2 on Hugging Face to access the models and dataset.



