What actually arrived on November 16

Read the release in order and the structure is plain. The model is not the news. Nemotron 3.5 Super VL is a 120-billion-parameter vision-language model with 12 billion active parameters, and its model card dates its Hugging Face release to October 15, a month before this announcement (card details as read by Unite.AI). NVIDIA is not hiding that; the release calls Nemotron for Science a bringing-together of applications, data and recipes around that model.

The new material is LitmusBench v0.2 and the training method around it. Nearly 50,000 chemistry questions is a substantial open dataset, and its design matters more than its size. The questions are built from molecular properties and experimental drug-discovery data, so a model's answer can be scored against a measured outcome rather than against another model's preference. That is the thread running through the whole package: pick tasks where a measured outcome can mark the work.

The recipe is the product

The accompanying technical guide makes the mechanism concrete. Experimental records in a spreadsheet become question-answer pairs. The measured result is the verifier. A model predicting what an enzyme does to a molecule can be scored on structural similarity to the real metabolite, on a zero-to-one scale, and trained against that score. Unite.AI reports that NVIDIA's guide gives a 12% improvement after supervised fine-tuning in that example and a further 17% after reinforcement learning, at a 16,384-token output budget.

That approach has a clear strength and a clear limit. The strength is that the reward signal comes from a measurement, not from another model's preference. The limit is that the example uses measured outcomes to construct a task-specific scoring signal, and that signal's usefulness depends on its accuracy and relevance to the research objective. Molecular fingerprints, simulation software and laboratory feedback can supply such signals; much of open science cannot, yet. The package is therefore best read as a method for the subset of science that can be auto-marked, which is a real subset and a growing one, but not science in general.

Read the partner numbers as NVIDIA's numbers

Every headline figure in the release comes from NVIDIA or from a partner quoted by NVIDIA. Lila Sciences at 61% on quantum-dot protocols against GPT-6 Astra at 54%, at a tenth to a twentieth of the inference cost. Therna at nearly double the accuracy of frontier models in predicting protein-expression ranges, with scientists preferring its explanations over the base model's in 96% of comparisons. KTH at 81% of its test cases for broken GROMACS setups repaired, with reasoning enabled, in one overnight run on a single DGX Spark.

Some of these have a built-in check that deserves credit: KTH's fixes were verified by GROMACS itself, not by the agent's own judgment. That is the right pattern, though it is a task-specific check, not independent replication. The release cites no independent replication for any of the figures, the task selection is the partners' and NVIDIA's, and a bounded task result does not establish that the same model will perform on a neighboring task in the same laboratory. Treat the numbers as evidence that post-training this open model on verifiable tasks can move specific scores. Do not treat them as a league table.

What a research team can do with it today

The practical offer is concrete. The model and the dataset are on Hugging Face now. The model card targets a 128GB DGX Spark, so a team can run it on infrastructure it controls, which matters for unpublished compounds and patient-adjacent data that cannot leave the building; teams should still validate the documented configuration and data handling before using sensitive material. The dataset repository reportedly lists CC-BY-SA-4.0 and OpenMDW-1.1; check which terms apply to the intended material and use before commercial work.

If your work has a scorer, a simulator, an assay, a solver, this release gives you an open model, a chemistry dataset and a worked recipe for training against that scorer. If your work has no verifier, you have downloaded a month-old model with a new press release.

Sources

  1. [1] NVIDIA, "NVIDIA Introduces Nemotron for Science to Accelerate Discoveries With Agents", press release via GlobeNewswire, 16 November 2026 (full text read at Investing News Network)Read source
  2. [2] Unite.AI, "NVIDIA Launches Nemotron for Science and Open Chemistry Dataset", 16 November 2026 (model-card, dataset-repository and technical-guide details)Read source
  3. [3] NVIDIA Nemotron 3.5 Super VL model page, Hugging FaceRead source