Nvidia Puts AI Agents in the Lab With Nemotron for Science
Nvidia unveiled Nemotron for Science at SC26 on November 16, 2026, pairing the multimodal Nemotron 3.5 Super VL model with an open chemistry dataset of nearly 50,000 reasoning questions. Lila Sciences hit 61% accuracy on quantum-dot synthesis, and Nvidia says its agents run at a fraction of rival inference costs.
- 3 sources report it: cryptobriefing.com, xenospectrum.com, blockchain.news.
- The source published it on Nov 16, 2026 at 4:16 PM.
Nothing open. Every check passed.
Nvidia dropped Nemotron for Science on November 16, 2026, at the SC26 conference — a package built to put specialized AI agents to work in the lab instead of the inbox. The bundle pairs a new multimodal model, Nemotron 3.5 Super VL, with an open chemistry dataset of nearly 50,000 reasoning questions. Nvidia's pitch: researchers can build domain-specific agents that run more accurately and cheaper than general-purpose rivals.
The model handles more than one kind of input, so it isn't stuck reading plain text. The chemistry dataset frames its questions as reasoning problems rather than simple fact lookups. Nvidia is aiming the whole kit at four fields: chemistry, life sciences, protein design and materials science. Both the models and the benchmarks are up on Hugging Face for public use.
The early numbers are loud. Lila Sciences hit 61% accuracy on quantum-dot synthesis using Nemotron, while GPT-6 Astra scored 54% on the same task. Inference costs for the Nemotron-based agents were reported at 10-20 times lower than competing models like GPT-6 Astra. A second showcase came from the KTH Royal Institute, where an agent fixed 81% of broken GROMACS setups in a single overnight run.
Institutions are already circling. Active users include the US Department of Energy's Argonne National Laboratory and Japan's RIKEN. The initiative extends Nvidia's habit of shipping open models alongside hefty datasets and post-training guidance, and it builds on earlier efforts like BioNeMo, the company's life sciences project. It also plugs into Nvidia's CUDA-X accelerated scientific computing suite, which is meant to make deployment work across a range of computing environments.
It all lands inside Nvidia's wider agentic AI push — systems that chain actions toward a goal rather than answering one prompt. Labs with tight budgets and long queues stand to gain most if the reported cost advantage holds across real workloads. Still, the quantum-dot win is one task, and one benchmark doesn't settle which model is better for science at large. And fixing 81% of broken GROMACS setups leaves the rest needing a human.
TopicsNvidia · Nemotron for Science · Nemotron 3.5 Super VL · Lila Sciences · GPT-6 Astra · Argonne National Laboratory · RIKEN · KTH Royal Institute
Written by AI from verified reporting.

