Nvidia launches Nemotron for Science to put AI agents in the lab
Nvidia has introduced Nemotron for Science, an initiative to accelerate scientific discovery with specialized AI agents, unveiled at the SC26 conference on November 16, 2026.
The package includes a new multimodal model, Nemotron 3.5 Super VL, and an open chemistry dataset of nearly 50,000 reasoning questions. Nvidia's pitch is that researchers can build domain-specific agents that are more accurate and cheaper to run than general-purpose rivals.
What Nvidia actually shipped
The core of the release is Nemotron 3.5 Super VL. A multimodal model can work with more than one kind of input, so it isn't limited to reading plain text. Alongside the model comes the chemistry dataset, with its nearly 50,000 questions framed as reasoning problems, not simple fact lookups.
Nvidia is aiming the tools at four areas: chemistry, life sciences, protein design and materials science. The models and benchmarks are available on Hugging Face for public use.
Early results
Lila Sciences reached 61% accuracy in quantum-dot synthesis using Nemotron. A competing model, GPT-6 Astra, scored 54% on the same work. Inference costs for the Nemotron-based agents were reported to be 10-20 times lower than competing models like GPT-6 Astra.
The second showcase came from the KTH Royal Institute. An agent built there fixed 81% of broken GROMACS setups in a single overnight run.
Institutional adoption
Institutional interest is already on the board. Active users include the US Department of Energy's Argonne National Laboratory and Japan's RIKEN. The release extends Nvidia's habit of releasing open models paired with substantial datasets and guidance for post-training.
Broader implications
The initiative builds on earlier work like BioNeMo, Nvidia's life sciences project, and plugs into the company's CUDA-X accelerated scientific computing suite. This integration is designed to make deployment possible across a range of computing environments.
The release also lands within Nvidia's broader push into agentic AI - systems that can take a sequence of actions toward a goal, not just answer a single prompt. For researchers with tight budgets, the reported 10-20 times lower inference cost could make AI-assisted research more accessible.