Money and business in the Middle East.

AI

Nvidia releases open science model and 50,000-question chemistry dataset

Nvidia launched Nemotron for Science, pairing an open model with chemistry data and post-training recipes for research agents

· Last verified: 17 Nov 2026 (Nvidia Newsroom)

Summary

  • Nvidia launched Nemotron for Science, pairing an open model with chemistry data and post-training recipes for research agents
  • Argonne, Jülich, RIKEN and KTH are already building discovery agents on the Nemotron 3.5 Super VL model
  • Labs can now run specialized science agents on hardware they control, from desktops to supercomputers

The latest

Nvidia on November 16, 2026 released Nemotron for Science, a package built around its open Nemotron 3.5 Super VL model, a new chemistry dataset and post-training recipes aimed at researchers building discovery agents. The company said the model has been optimized to run locally on a 128GB DGX Spark system as well as in data centers, keeping scientific workflows on infrastructure organizations control.

Details

  • The dataset: LitmusBench v0.2 is an open dataset of nearly 50,000 questions for training and evaluating language models on chemistry reasoning, according to Nvidia. It draws on molecular properties and experimental drug discovery data, and is meant to improve predictions of how molecules behave, including solubility and how the body processes them.
  • Availability: Nvidia said both Nemotron 3.5 Super VL and LitmusBench v0.2 are available now on Hugging Face. The company did not attach pricing or licensing conditions to the release, and did not name a date for further additions to the package.
  • The pitch: Ian Buck, Nvidia's vice president of hyperscale and high-performance computing, said the package 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."
  • Chemistry extraction: Terray Therapeutics post-trained Nemotron to pull reactants, products and reaction conditions out of scientific diagrams, surpassing previously published results on two benchmarks, according to Nvidia. On the RxnScribe test it recovered about 76% of annotated reactions under the benchmark's matching criteria.
  • Lab protocols: Lila Sciences used the model to design execution-ready laboratory protocols for its automated labs. Nvidia said post-training with laboratory feedback nearly doubled accuracy for making quantum dots to 61%, against 54% for GPT-6 Astra, at 10-20x lower inference cost.
  • Therapeutics: Therna worked with Nvidia to post-train Nemotron for agentic mRNA therapeutic design, nearly doubling accuracy in predicting protein-expression ranges over frontier models, the company said. Scientists preferred its explanations over those of the base model in 96% of comparisons.
  • National labs: The U.S. Department of Energy's Argonne National Laboratory is integrating Nemotron into its StructBioReasoner multi-agent system to design and evaluate protein-based therapeutic candidates for cancer and other diseases. Jülich Supercomputing Centre is folding the model into its own discovery agents.
  • Autonomous repair: KTH Royal Institute of Technology built an agent on Nemotron that repairs broken GROMACS molecular dynamics setups unaided. In one overnight run on a single DGX Spark system, the agent fixed 81% of test cases with reasoning enabled, and every fix was checked independently by GROMACS itself.
  • Japan: RIKEN has made Nemotron, including the 3.5 Super VL model, available on its RIKYU AI supercomputing platform through RiVault, its inference gateway, Nvidia said.

Background

Nemotron is Nvidia's family of open models, released with weights and post-training recipes rather than as a closed API. That design lets institutions adapt models to narrow tasks and run them on their own hardware, a requirement for labs handling sensitive experimental and patient-linked data.

Between the lines

The benchmark figures Nvidia published are self-reported or supplied by partner companies, and several compare against frontier models on narrow tasks: quantum dot synthesis, protein-expression prediction, reaction extraction. The cost claim is the sharper argument — Lila Sciences reporting 61% accuracy at 10-20x lower inference cost than a frontier rival reframes open weights on local hardware as an economic case, not only a data-control one.

What's next

Watch whether independent researchers reproduce the RxnScribe and quantum dot results on the public Hugging Face release, and whether Argonne's StructBioReasoner system produces named therapeutic candidates.

Source: Nvidia Newsroom

Nvidia releases open science model and 50,000-question chemistry dataset · INXEN