Decode News AI
简体繁體

Nvidia Launches Nemotron for Science With Open Chemistry Dataset

2026-11-16 12:35 EST · AI-compiled · 281 words · about 1 min · Views …
Photo: AI-generated illustration

Nvidia introduced its Nemotron for Science platform at the SC26 supercomputing conference on Nov. 16, 2026. The platform pairs an open multimodal model with nearly 50,000 chemistry reasoning questions. Nvidia says the aim is to speed up molecular research.

Open Chemistry Dataset Boosts Drug and Material Modeling

The release centers on the Nemotron 3.5 Super VL open model and LitmusBench v0.2. LitmusBench is an open dataset containing nearly 50,000 questions built from experimental drug discovery data. Researchers can use the package to train artificial intelligence agents to predict molecular solubility and how the human body processes compounds. Such agents are software systems that plan tasks and run simulations.

In a drug-metabolism test, supervised fine-tuning raised accuracy by 12%. Reinforcement learning post-training lifted it another 17%. Ian Buck is Nvidia vice president of hyperscale and high-performance computing.

Biotech Firms and National Labs Deploy Early Workflows

Life sciences firm Lila Sciences used laboratory feedback to achieve 61% accuracy in synthesizing quantum dots. That beat GPT-6 Astra at 54%, with 10 to 20 times lower inference costs. Terray Therapeutics adapted the model to recover about 76% of annotated chemical reactions from scientific diagrams on the RxnScribe test.

KTH Royal Institute of Technology also built an agent that fixed 81% of broken molecular dynamics setups overnight on a single DGX Spark system. Argonne National Laboratory and Japan's RIKEN integrated the model. Both the model and dataset are available on Hugging Face.

Sources

Found a factual error or typo? Tell us

Follow Decode News AI on Telegram

Live updates: every new story is pushed the moment it publishes — dozens a day, no digest delay.

@decodenewsai · Subscribe now

Free · unsubscribe anytime

Comments