BlockBeats news, November 17 — According to an official NVIDIA press release, the company has launched NVIDIA Nemotron for Science, integrating the open-source model Nemotron 3.5 Super VL, a new open chemistry dataset, and post-training recipes to help researchers build specialized agents for scientific discovery. Among them, the open dataset LitmusBench v0.2 contains nearly 50,000 chemistry reasoning questions, sourced from molecular property and experimental drug discovery data, and can be used to train and evaluate models' predictions of molecular behaviors such as solubility and human metabolism. The model and dataset are now available on Hugging Face and have been optimized for local operation on 128GB NVIDIA DGX Spark systems.
Life sciences companies Lila Sciences, Terray Therapeutics, and Therna are using and customizing the model. After post-training, Terray uses it to extract reactants, products, and reaction conditions from scientific charts, recovering approximately 76% of annotated reactions in RxnScribe tests according to the benchmark's matching criteria. Lila Sciences uses it to design executable experimental protocols, stating that it achieves performance comparable to frontier models without additional fine-tuning; after post-training with laboratory feedback, the accuracy of preparing quantum dots nearly doubled to 61%, surpassing GPT-6 Astra's 54%, with inference costs approximately one-tenth to one-twentieth of the latter. Therna, in collaboration with NVIDIA, post-trained the model for mRNA therapeutic design, stating that the accuracy of predicting protein expression ranges nearly doubled compared to frontier models, and scientists preferred its explanations in 96% of comparisons.
The U.S. Department of Energy's Argonne National Laboratory is integrating it into the StructBioReasoner multi-agent system for designing and evaluating protein candidate therapeutics for diseases such as cancer, and the Jülich Supercomputing Centre is also adopting the model. KTH Royal Institute of Technology has built an agent based on the model that can autonomously repair broken GROMACS molecular dynamics setups; after running overnight on a single DGX Spark with reasoning enabled, it repaired 81% of test cases, all independently verified by GROMACS. RIKEN has made models including Nemotron 3.5 Super VL available on the RIKYU AI supercomputing platform through the inference gateway RiVault. Ian Buck, NVIDIA's Vice President of Hyperscale and High-Performance Computing, said the solution provides researchers with open-source models, scientific data, and post-training recipes to build specialized agents capable of reasoning about scientific problems and invoking CUDA-X accelerated tools.
