NVIDIA has launched Nemotron for Science, a package meant to help researchers build AI agents for chemistry, biology and materials work on an open model they can run on their own hardware. Announced on 16 November 2026, it combines the Nemotron 3.5 Super VL open model, a new open chemistry dataset called LitmusBench v0.2, scientific evaluations and post-training recipes. NVIDIA says Argonne National Laboratory, Jülich Supercomputing Centre, RIKEN and companies including Lila Sciences, Terray Therapeutics and Therna are already using it.
Quick Take
- What it is: a science toolkit built around NVIDIA’s open Nemotron 3.5 Super VL vision-language model, plus data and recipes for post-training.
- New dataset: LitmusBench v0.2, nearly 50,000 chemistry reasoning questions for training and evaluation, available on Hugging Face.
- Headline claim: Lila Sciences says post-training with lab feedback nearly doubled accuracy for making quantum dots to 61%, above GPT-6 Astra’s 54% at 10 to 20 times lower inference cost, according to NVIDIA.
- Where it runs: from a 128GB NVIDIA DGX Spark desktop up to supercomputing centers.
- Who uses it: Argonne, Jülich, KTH, RIKEN, Lila Sciences, Terray Therapeutics and Therna.
In plain English
NVIDIA is giving scientists a free AI model, a big set of chemistry practice questions and instructions for teaching the model their own specialty. The goal is AI assistants that can read papers and diagrams, plan experiments and run scientific software, while the data stays on the lab’s own machines. NVIDIA’s partners report that a specialized version can beat a much larger frontier model on narrow tasks for far less money. Those results come from NVIDIA and its partners, not independent tests.
What happened
In its announcement, NVIDIA says Nemotron for Science “brings together the Nemotron 3.5 Super VL model, scientific evaluations, a new open chemistry dataset and recipes to help researchers build domain-specific science agents.” Researchers can post-train the model with their own data and plug it into agent workflows that “interpret scientific information, generate code and call computational tools,” using NVIDIA’s CUDA-X accelerated libraries for the heavy computation.
“Nemotron for Science gives researchers open models, scientific data and post-training recipes to build specialized agents that reason through scientific problems,” said Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing.
What’s in it
LitmusBench v0.2 has nearly 50,000 questions for training and evaluating models on chemistry reasoning. NVIDIA says it draws on molecular properties and experimental drug discovery data, to help models predict how molecules behave, including solubility and how the body processes them.
Partner results, as reported by NVIDIA:
- Terray Therapeutics post-trained Nemotron to extract reactants, products and conditions from chemistry diagrams, recovering about 76% of annotated reactions on the RxnScribe test.
- Lila Sciences used it to design lab protocols for automated labs. Post-training with lab feedback nearly doubled accuracy for quantum dot synthesis to 61%, compared with 54% for GPT-6 Astra.
- Therna post-trained it for mRNA therapeutic design and says it nearly doubled accuracy in predicting protein-expression ranges over frontier models.
- KTH Royal Institute of Technology built an agent that repairs broken GROMACS molecular dynamics setups. In one overnight run on a single DGX Spark, it fixed 81% of test cases, each checked by GROMACS itself.
- Argonne is integrating Nemotron into its StructBioReasoner multi-agent system for protein-based therapeutics, and RIKEN has made the models available on its RIKYU platform.
Why it matters
For researchers, the pitch is control. Labs with sensitive or unpublished data often cannot send it to a cloud model. An open model that runs on a DGX Spark and can be specialized for one task is a different trade from renting a frontier model, and NVIDIA’s partner numbers suggest a small, tuned model can win on narrow problems.
For the AI science race, this is NVIDIA’s answer to the frontier labs’ science push, including OpenAI for Science and the recent GPT-6 Astra lattice proof with physicist Nathaniel Craig. NVIDIA sells the hardware either way, but open models keep researchers building on its stack.
For companies in pharma and materials, the recipes matter as much as the weights. The value is in post-training on proprietary lab feedback, which is exactly the data these companies will not share.
What’s still open
- Independent replication of the partner results, which are reported by NVIDIA and measured on partner-chosen tasks.
- The license terms for Nemotron 3.5 Super VL and LitmusBench v0.2 for commercial use.
- How the model compares with other open science models on shared benchmarks.
- How much compute and lab data a typical team needs to reach the reported gains.
Sources
- NVIDIA Newsroom, “NVIDIA Introduces Nemotron for Science to Accelerate Discoveries With Agents,” 16 November 2026
i10x Perspective
The most useful claim in this launch is not a benchmark score, it is the shape of the result: a mid-sized open model, tuned on a lab’s own feedback, beating a frontier model on that lab’s task at a fraction of the cost. If that pattern holds, scientific AI will look less like one giant model and more like many specialized ones, owned by the labs that trained them. For more on open releases, follow our Models section.






