Overview
- NVIDIA unveiled Nemotron for Science on Monday, November 16, 2026, and published Nemotron 3.5 Super VL plus the LitmusBench v0.2 chemistry dataset on Hugging Face for public use and local deployment.
- The bundle includes an open multimodal model, post‑training recipes aimed at chemistry, life sciences, protein design and materials science, and a near‑50,000 question dataset framed as reasoning tasks.
- Company and partner demonstrations reported initial wins: Lila Sciences reached 61% accuracy on a quantum‑dot synthesis task versus 54% for a rival model, and a KTH agent repaired 81% of broken GROMACS setups in one overnight run.
- NVIDIA said expanded CUDA‑X toolkits deliver large speedups in scientific workloads — examples include reducing Event Horizon Telescope correlation to eight minutes, up to 3× molecular simulation speedups, and a reported 900× cell‑graph speedup at Memorial Sloan Kettering — and claimed Nemotron inference can cost 10–20× less in partner tests.
- Alongside software, NVIDIA announced future hardware and collaborations, including the Rubin HPC GPU, expanded Vera CPUs for upcoming supercomputers, and a quantum co‑design effort with IBM and Fermilab, while noting partners and public releases enable outside verification of the performance claims.