NVIDIA's Nemotron for Science Brings Open AI Agents to the Chemistry Lab

Photo: Hippopx (public domain). Illustrative image.
What happened. NVIDIA introduced Nemotron for Science at SC26 on 16 November 2026, a bundle for building AI agents that do science: an open model, an open dataset and post-training recipes, all available now on Hugging Face. The model is Nemotron 3.5 Super VL — 120 billion parameters in total, 12 billion active, with up to 1 million tokens of context — optimised to run locally on a 128 GB DGX Spark. The dataset is LitmusBench v0.2, an open chemistry reasoning set of about 50,000 questions (47,118 training and 2,056 test, drawn from ChEMBL v36, ExpansionRx and Biogen). The recipes show how to post-train the model into a science agent.
Who is using it. NVIDIA names Lila Sciences, Terray Therapeutics and Therna as users, with Argonne National Lab integrating the model into StructBioReasoner and RIKEN serving it through RiVault. The most concrete example comes from KTH, where an agent repaired 81% of broken GROMACS setups overnight on a single DGX Spark — the kind of overnight, unattended repair work that is the whole point of a science agent. The bundle sits alongside Rubin HPC, NVIDIA’s new 50,000-core GPU for agent-driven simulation, announced at the same conference.
The claims need labelling. Every performance figure in the release is NVIDIA’s own, with no independent verification published. Lila’s quantum-dot accuracy is given as 61%, with the release comparing it against “GPT-6 Astra’s 54%” — a comparison that must be framed strictly as NVIDIA’s claim, because GPT-6 Astra is itself an unconfirmed model name that has only appeared inside OpenAI’s own announcements. Terray is cited at 76% on RxnScribe, Therna at “96% explanation preference”, and the bundle is said to have “led 7 of 11 science benchmarks”. None of these has an independent check yet. The open parts — the model weights, the dataset, the recipes — are verifiable by anyone who downloads them. The leaderboard claims are not, yet.
Why it matters. The interesting part of Nemotron for Science is not the model but the shape of the bundle: model plus benchmark plus recipes, all open, all runnable on a desktop. That is a bet that science agents will be built by labs, not just bought from labs — the same logic behind the CRAM energy model, a 7,700-parameter reasoning system published as an open paper last week. Chemistry is the right first target: it has structured data, clear right answers, and expensive experiments, so an agent that can propose and check a molecule is worth real money. Whether Nemotron 3.5 Super VL is actually the best open model for that job will be settled by the community now that it can run it — which is precisely why NVIDIA opened it.
Sources
- NVIDIA / GlobeNewswire — “NVIDIA Introduces Nemotron for Science to Accelerate Discoveries with Agents” (16 November 2026): https://www.globenewswire.com/news-release/2026/11/16/3403880/0/en/nvidia-introduces-nemotron-for-science-to-accelerate-discoveries-with-agents.html
- FreshSynapse — “NVIDIA Rubin HPC Packs 50,000+ CUDA Cores”: https://freshsynapse.com/articles/nvidia-rubin-hpc-50000-cuda-cores-2026-11-17/
- FreshSynapse — “CRAM: A 7,700-Parameter Energy Model That Trains 1,000x Faster”: https://freshsynapse.com/articles/cram-energy-model-reasoning-1000x-faster-2026-11-16/