Technology / Science

Add as a preferred source on Google

NVIDIA Pairs Nemotron AI Model With 49,174 Chemistry Questions

The Hugging Face dataset card lists 47,118 training questions and 2,056 test questions across 126 properties and 12 assay endpoints.

Updated 2 min read
A conference attendee works at a compact AI workstation during SC26 in Chicago. / TokenPost.ai
A conference attendee works at a compact AI workstation during SC26 in Chicago. / TokenPost.ai

NVIDIA introduced a science-focused AI package at SC26 in Chicago on Nov. 16, pairing its Nemotron model with chemistry data and training methods researchers can use to adapt agents for scientific tasks.

The package centers on Nemotron 3.5 Super VL and Litmus-Bench v0.2, alongside post-training recipes for research agents. NVIDIA lists both the model and dataset on Hugging Face and says the model can run locally on an NVIDIA DGX Spark system with 128GB of memory.

The dataset card lists 47,118 training questions and 2,056 test questions, 49,174 in all. It spans 126 properties and 12 assay endpoints, drawing on ChEMBL, ExpansionRx and Biogen; some questions also use descriptors produced with RDKit.

NVIDIA's technical guide uses experimental records to check model answers before supervised fine-tuning and reinforcement learning. For a drug-metabolism task, NVIDIA reported a 12% accuracy gain after fine-tuning and a further 17% after reinforcement learning, using a 16,384-token output budget. The evaluation compared predicted metabolites with experimental records using molecular fingerprints and Tanimoto similarity.

The release follows a CUDA-focused Nemotron adaptation previously reported by EconoTimes, in which Mindbeam said it post-trained Nemotron 3.5 Super VL for GPU kernel generation.

NVIDIA also cited partner test results. It said Terray Therapeutics' post-trained Nemotron model recovered about 76% of annotated reactions under the RxnScribe benchmark's matching criteria. For quantum-dot synthesis, NVIDIA said Lila Sciences' lab-feedback post-trained Nemotron model reached 61% accuracy, compared with 54% for GPT-6 Astra. The announcement did not provide test counts or full evaluation methods for either comparison.

NVIDIA said Argonne National Laboratory is integrating Nemotron into StructBioReasoner. An Argonne case study describes an earlier version of the system using GPT-OSS-120B as its reasoning agent and says experimental validation of designed candidates is still needed. NVIDIA also said RIKEN has made Nemotron models available through RiVault on its RIKYU AI supercomputing platform.

Related articles

View all