Particle.news

Nvidia Unveils Nemotron for Science With Open Multimodal Model and Chemistry Benchmarks

The company says the toolkit could make AI-driven experiment design and lab software debugging far cheaper to run on local GPUs.

Overview

  • Nvidia launched Nemotron for Science on Monday with Nemotron 3.5 Super VL, a multimodal model, and an open chemistry reasoning dataset of nearly 50,000 questions posted on Hugging Face.
  • The public package includes LitmusBench v0.2 benchmarks plus post‑training recipes to help researchers build agentic workflows that use multiple input types such as text and diagrams.
  • Early demonstrations reported by Nvidia showed task wins — Lila Sciences measured 61% accuracy on a quantum‑dot synthesis task versus 54% for GPT‑6 Astra and KTH said an agent fixed 81% of broken GROMACS setups overnight — but these results are company‑reported and need independent validation.
  • Nvidia integrated the release with its BioNeMo efforts and CUDA‑X accelerated toolkits so users including Argonne and RIKEN can run models on local clusters or DGX systems to reduce latency and control data workflows.
  • If the claimed 10–20x lower inference costs hold up in real labs, smaller groups could afford agentic AI for experiment planning and simulation, but generality limits and persistent failure modes mean human oversight will remain essential.