Latest

Artificial Intelligence

NVIDIA opens Nemotron for Science to scientific agents

NVIDIA releases the Nemotron 3.5 Super VL model, a chemistry question bank of nearly 50,000 cases and recipes for laboratories to train their own agents.

Translated from Spanish by an automated system. The Spanish original prevails. Read the Spanish original

NVIDIA opens Nemotron for Science to scientific agents
NVIDIA opens Nemotron for Science to scientific agents

A research agent that proposes experiments needs more than a large model: it needs chemistry data, a way to assess whether it reasons well and a place to run without sending results to an external provider every time it runs a test. NVIDIA presented on November 16, 2026 Nemotron for Science, a package that brings those three pieces together around its open model Nemotron 3.5 Super VL.

The set includes the model, an evaluation bank with nearly 50,000 chemistry questions called LitmusBench v0.2 and post-training recipes to adapt it to laboratory tasks. The proposal is that a research group can build its own specialized agent without starting from scratch and without depending on a commercial interface for each test, something that until now required a dedicated engineering team.

The package also includes fine-tuning recipes, and the announcement fits a strategy NVIDIA has repeated this year: sell the hardware and give away the software that makes it necessary, so that every laboratory that adopts an agent becomes a customer for graphics cards. If laboratories train scientific agents, they need graphics cards to train and to infer, and the manufacturer ensures the ecosystem grows on its platform.

An all-in-one package

Nemotron 3.5 Super VL is an open model with multimodal capabilities, meaning it processes text and images in the same conversation. That feature matters in chemistry, where data arrives in the form of molecular structures, spectra and tables that a text-only model cannot read directly.

LitmusBench v0.2 is the most useful component for a researcher. It is a set of nearly 50,000 questions on molecular properties and drug discovery data, available on Hugging Face, which serves for training and for evaluation. The difference between a question bank and a demonstration is that the former allows measuring whether the model improves or only appears to improve.

The post-training recipes complete the package. NVIDIA describes how to adapt the model to tasks such as designing proteins or reasoning about reactions, with examples of supervised fine-tuning and reinforcement. A laboratory with its own data can follow those recipes and obtain a model specialized in its domain without publishing its data.

The fourth element is infrastructure. NVIDIA places training on DGX Spark, its desktop machine, and in data centers, with the promise that the workflow works the same in both environments. For a small group, that continuity is the difference between testing an idea and setting up a cluster.

Chemistry as a test bed

Chemistry has a property that makes it attractive for agents: its results can be verified. A prediction about the solubility of a molecule is checked against an experiment, and that binary answer allows training with reinforcement without depending on a human to score each response. That characteristic is what has made molecular design one of the first domains where agents actually work.

It is also a domain where mistakes are costly. A model that proposes a dangerous synthesis or confuses one molecule with another can cause a laboratory accident, and that is why evaluations have to include adverse cases and not only exam questions. LitmusBench includes questions on properties and on processing in the human body, according to the manufacturer's description.

The third reason is data scale. The chemistry literature is enormous and is structured in patents, articles and databases, and that volume allows training specialized models without depending on a single company's proprietary data. An academic group with access to the same sources can compete with a pharmaceutical laboratory on the model side, though not on the experiments side.

The fourth is economic. Discovering a molecule costs years and millions of dollars, and any tool that reduces the number of experiments needed has a return that can be measured. That return justifies investing in own infrastructure, which is exactly what NVIDIA sells.

To those three reasons is added a fourth that explains the interest of large laboratories: the possibility of training with their own data without publishing it and without the provider using it to train other models.

The package arrives at a time when laboratories are debating whether to train their own models or rent others' interfaces, and that doubt is what the manufacturer wants to resolve with its offering.

Who is using it?

NVIDIA cites Argonne National Laboratory, one of the research centers of the United States Department of Energy, as a user of Nemotron to build agents that accelerate scientific discoveries and then share the results with the community. It also mentions Lila Sciences, Terray Therapeutics and Therna, three companies working in drug discovery that adapt the model to their workflows.

Those names matter because they place the announcement outside the realm of demonstration. A national laboratory and three privately funded companies are using the package in production or in advanced pilots, and that use generates data that returns to the ecosystem in the form of recipes and evaluations.

The fine print is that none of the published cases includes comparable results. There are no figures on how much a discovery workflow has improved or how many experiments have been saved, and without those figures the announcement describes a capability and not a result.

The difference between capability and result is the same as the difference between a model that passes a chemistry exam and an agent that discovers a molecule that works outside the laboratory.

What a laboratory needs to test it

The first requirement is having its own data in a usable format. A group that has spent years accumulating paper laboratory notebooks cannot train anything until it digitizes and structures that information, and that work usually takes longer than fine-tuning the model.

The second is knowing how to evaluate. LitmusBench provides a starting point, and each domain needs its own set of questions with verified answers. A model that gets general chemistry right can fail on the specific reaction the group cares about, and the only way to know is to measure it with its own cases.

The third is infrastructure. The open model can be run locally, and that requires cards with enough memory for the size of the model. NVIDIA sells that infrastructure, so its recommendation is not neutral, although the alternative of using a commercial interface also has a cost.

The fourth is data governance. An agent that reads patents and internal results needs permissions and traceability, and in a pharmaceutical laboratory that layer is as important as the model.

The open model as a commercial strategy

NVIDIA does not make money from the model, it makes money from what is needed to use it. That logic explains why it publishes open weights, evaluation banks and recipes: each piece lowers the barrier for a laboratory to set up its own system on the company's hardware. Free software here is the instruction manual for a machine that is sold separately.

The calculation carries a risk. If open models improve enough, customers can run them on competitors' hardware, and the manufacturer loses the advantage it sought. NVIDIA offsets this with an ecosystem of tools and optimization agreements that make its platform the most comfortable to start with.

For the scientific sector, the result is a race for reproducibility. An agent that discovers a molecule has to be able to explain how it got there, and a closed model does not always allow that. NVIDIA's bet is that laboratories will prefer a model they can inspect, even if it costs them more effort to get it running.

Science as the next customer for agents

In two years agents have gone from writing emails to proposing experiments, and that leap requires verifiable data and rigorous evaluations. Chemistry is the first domain where the three things coincide, and that is why announcements pile up in that field.

What comes next depends on how many laboratories manage to obtain publishable results with these tools. A software package does not change science, but it does change the time it takes to rule out a hypothesis, which is the part of the work nobody boasts about doing.

For the reader who does research, the useful question is not whether the model is good, it is whether their data is in a condition to train it and whether they have a set of cases to check it. The rest is bought. NVIDIA's technical blog details how the model is fine-tuned and what it expects from each task. The comparison with Meta's cheap model for agents helps place the price: one competes on cost per token and the other on control of the data.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
Fed warns of risk of mass deposit withdrawal due to AI

Keep reading