Last updated: November 17, 2026
NVIDIA Nemotron for Science, announced November 16, 2026, packages the open Nemotron 3.5 Super VL model with a new chemistry dataset, LitmusBench v0.2, scientific evaluations and post-training recipes. Researchers can download the model and data free from Hugging Face, run them on a DGX Spark or data center GPUs, and adapt the model to their own scientific tasks.
Contents
- What NVIDIA released
- What is in Nemotron for Science
- How to run the model
- How to post-train for your science task
- Reported results
- Compared with the October release
- Limitations
- Frequently asked questions
What NVIDIA released
NVIDIA introduced Nemotron for Science on November 16, 2026, during its SC26 special address. The press release describes it as a combination of scientific applications of the Nemotron 3.5 Super VL open model, a new open chemistry dataset and post-training recipes that help researchers build specialized agents for scientific discovery.
The idea is to start from a capable open model, then post-train it on your own experimental data so an agent can reason through problems and call computational tools.
What is in Nemotron for Science
| Component | What it is (per NVIDIA) |
|---|---|
| Nemotron 3.5 Super VL | Open vision-language model, 120B total and 12B active parameters, up to 1M-token context, text, image and video input |
| LitmusBench v0.2 | Open chemistry reasoning dataset: 47,118 training and 2,056 test questions, 126 properties and 12 assay endpoints, about 64 MB |
| Post-training recipe | Builds a reinforcement learning environment from experimental data with NeMo Gym and trains with NeMo RL |
| Scientific evaluations | Benchmarks used to measure the model before and after adaptation |
| License | Model: OpenMDW 1.1. Dataset: CC BY-SA 4.0 for ChEMBL-derived questions, OpenMDW 1.1 for the rest |
| Price | No price listed; free download, you supply the compute |
The LitmusBench v0.2 dataset card says the questions come from the ChEMBL, ExpansionRx and Biogen datasets, some use RDKit descriptors, and every question needs a short answer. It is released as part of NeMo Gym. NVIDIA also writes the name as Litmus-Bench v0.2 or Litmus-Bench V2.0; all refer to the same dataset.
How to run the model
The model ships in two main formats. The BF16 card is the full-precision release for customization and evaluation and lists 4x H100 or 2x GB200 for deployment. The NVFP4 card is the optimized release and lists 1x DGX Spark, 1x RTX PRO 6000, 1x H200, 2x H100 or 1x GB200.
- Choose the NVFP4 weights for a DGX Spark or a single large GPU, or the BF16 weights if you plan to post-train.
- Pull a vLLM OpenAI-compatible Docker image. The BF16 card names
vllm/vllm-openai:v0.32.0or a specific nightly build. - Set
HF_TOKENif Hugging Face asks for authentication, then runvllm servewith the model ID,--trust-remote-code, thenemotron_v3reasoning parser and theqwen3_codertool-call parser shown on the card. - Check the server with
curl --fail http://localhost:8000/v1/models. The examples use the API model namenemotron-3.5-super-vlon port 8000. - Call it with any OpenAI-compatible client, and switch reasoning on or off with
enable_thinkingin the chat template.
On DGX Spark, the NVFP4 card recommends the LiLiCorr speculative drafter in all scenarios. NVIDIA also points DGX Spark and RTX users to llama.cpp and Unsloth for local inference.
For background on the serving stack, see our vLLM 0.33 guide and our llama.cpp 0.9.0 guide.
How to post-train for your science task
The NVIDIA technical blog post walks through a drug-metabolism example: given a parent compound and an enzyme, the model predicts the resulting metabolite.
- Start with your records. Use a CSV of experimental data, such as SMILES strings, sequences or reaction conditions paired with measured outcomes.
- Convert rows to prompts. Turn each row into a JSONL prompt with the ground-truth answer stored alongside it for NeMo Gym.
- Write a verifier. The example scores a submitted molecule against the reference with Morgan-fingerprint Tanimoto similarity, from 0 to 1. Richer verifiers, such as simulators or a live automated lab, can be wired in as tool calls.
- Optionally warm-start. Run a stronger teacher model on the same environment, keep only correct reasoning traces and use them for supervised fine-tuning.
- Run RL. Train with group relative policy optimization (GRPO) and LoRA adapters in NeMo RL, using the verifier score as the reward.
- Evaluate on held-out data to confirm the trained model beats the starting model.
Because the reward comes straight from experimental data, this approach, which NVIDIA calls reinforcement learning with verifiable rewards, needs no human preference labels. NVIDIA advises starting with one narrow task. The example supervised fine-tuning command runs on 4 nodes with 4 GPUs each, so plan for multi-GPU capacity even if inference fits on one machine.
Reported results
These figures come from NVIDIA and its partners and have not been independently verified here.
- Drug metabolism (NVIDIA): at a 16,384-token output budget, supervised fine-tuning improved accuracy by 12% and RL post-training added a further 17%. The distillation step kept 1,123 high-quality traces from 1,934 training prompts at a Tanimoto threshold of 0.7 or higher.
- Lila Sciences: post-training with lab feedback nearly doubled accuracy for making quantum dots, to 61%, compared with 54% for GPT-6 Astra, at 10 to 20 times lower inference cost.
- Terray Therapeutics: recovered about 76% of annotated reactions on the RxnScribe test when extracting reactions from diagrams.
- KTH Royal Institute of Technology: an agent fixed 81% of broken GROMACS test cases in one overnight run on a single DGX Spark.
Compared with the October release
Nemotron 3.5 Super VL itself is not new; the model cards list a Hugging Face release date of October 15, 2026, which our Nemotron 3.5 guide covers. What changed on November 16 is the science-specific tooling around it.
| Question | Nemotron 3.5 Super VL (October 15) | Nemotron for Science (November 16) |
|---|---|---|
| Model weights | BF16 and NVFP4 on Hugging Face | Same model |
| Chemistry dataset | Not included | LitmusBench v0.2 (v0.1 was the previous version) |
| Post-training guidance | General customization | Step-by-step RL recipe with NeMo Gym and NeMo RL |
| Target user | Developers building general agents | Researchers building science agents |
If you want a smaller open model to fine-tune instead, our Inkling 2 Small open weights guide covers another recent option.
Limitations
- Hardware. Inference fits a DGX Spark with NVFP4, but the example training run uses 16 GPUs.
- Chemistry focus. The new dataset covers chemistry and molecular properties; other fields need your own data and verifier.
- Share-alike terms. ChEMBL-derived questions are under CC BY-SA 4.0, which matters if you redistribute derived data.
- No hosted price. NVIDIA lists no managed per-token pricing on these pages.
Frequently asked questions
What is NVIDIA Nemotron for Science?
It is a collection NVIDIA announced on November 16, 2026, that combines the open Nemotron 3.5 Super VL model, scientific evaluations, the LitmusBench v0.2 chemistry dataset and post-training recipes, so researchers can build agents for chemistry, life sciences and other scientific tasks.
Is Nemotron for Science free to use?
The model and dataset are free downloads on Hugging Face. The model is released under the OpenMDW License Agreement 1.1 and is marked ready for commercial or non-commercial use. NVIDIA does not list a price, but you pay for your own GPUs or cloud compute.
What is LitmusBench v0.2?
It is an open chemistry reasoning dataset with 47,118 training and 2,056 test short-answer questions built from ChEMBL, ExpansionRx and Biogen data. It is meant for reinforcement learning and evaluation with NVIDIA NeMo Gym.
Can I run Nemotron for Science on a DGX Spark?
Yes. NVIDIA says Nemotron 3.5 Super VL was optimized to run locally on a 128GB DGX Spark. The NVFP4 model card lists a single DGX Spark as a supported vLLM, SGLang and TensorRT-LLM deployment.
Is Nemotron for Science a new model?
No. It is built on Nemotron 3.5 Super VL, which was released on Hugging Face on October 15, 2026. The new parts are the chemistry dataset, the scientific evaluations and the post-training recipes.
Sources
- NVIDIA Newsroom: NVIDIA Introduces Nemotron for Science to Accelerate Discoveries With Agents (November 16, 2026)
- NVIDIA Technical Blog: Post-Training NVIDIA Nemotron to Increase Accuracy in Your Scientific Domain (November 16, 2026)
- Hugging Face: nvidia/Nemotron-RL-litmus-bench-v0.2 dataset card
- Hugging Face: NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 model card
- Hugging Face: NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 model card
Image: Nemotron for Science announcement image by NVIDIA, from the NVIDIA Newsroom, used for editorial purposes.