NVIDIA is combining open models designed for scientific research with expanded GPU-accelerated computing libraries to help researchers analyze complex data, run simulations, and speed up discoveries across fields ranging from astrophysics to drug development.
The company introduced NVIDIA Nemotron for Science, a package built around its Nemotron 3.5 Super VL model, a new open chemistry dataset and post-training recipes. It also expanded its CUDA-X libraries with tools for scientific workloads, including black-hole imaging, molecular simulations, materials research and single-cell biology.
The developments aim to help researchers build specialized agents that can reason through scientific problems and use computational tools to carry out research tasks.
A key addition is LitmusBench v0.2, an open dataset containing nearly 50,000 questions for training and evaluating models on chemistry reasoning. Drawing on molecular properties and experimental drug-discovery data, it helps researchers improve predictions of molecular behavior, including solubility and how the human body processes compounds.
NVIDIA’s CUDA-X expansion targets the computing bottlenecks that can slow down scientific work. Its cuPhoton toolkit, for example, includes modules that accelerate the processing of astronomical observations.
Using cuPhoton.xBeam, researchers can correlate data from the Event Horizon Telescope in eight minutes, compared with more than a month previously, NVIDIA said. The telescope links radio dishes around the world to create an Earth-sized observing instrument, generating petabytes of data that require extensive processing.
xLens module to extract insights while plasma simulations were running. Coupled with a locally running Nemotron-powered agent, the workflow could help researchers analyze results and steer simulations in real time.
Source link







