Pull to refresh
Logo
NVIDIA releases open chemistry dataset and tuning recipes for lab AI agents

NVIDIA releases open chemistry dataset and tuning recipes for lab AI agents

New Capabilities

Nemotron for Science pairs a 120-billion-parameter open model with a 49,000-question chemistry benchmark and guides for fine-tuning on private lab data

Overview

Updated 5 hours ago

NVIDIA's Nemotron for Science bundles three things: an open 120-billion-parameter vision-language model, a chemistry question set of nearly 50,000 items, and recipes for tuning both on a lab's own data. The target is agents that read a reaction diagram, plan an experiment, or predict how a molecule behaves in the body.

The pitch is ownership. A lab that tunes an open model keeps the weights and the data. A lab that calls a frontier model through an API keeps neither. Three partner companies report that tuned Nemotron beat frontier models on narrow chemistry tasks at lower cost. Those results come from the companies themselves, on tasks they picked, which makes outside replication the thing to watch.

Why it matters

If labs can tune a free model on private data and beat a frontier API on narrow tasks, discovery work stops renting intelligence.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

49,174
LitmusBench v0.2 questions
47,118 training and 2,056 test questions on chemistry reasoning, built from ChEMBL, ExpansionRx and Biogen data.
120B
Parameters in Nemotron 3.5 Super VL
A 120B-A12B model, meaning 12 billion parameters active per token; runs locally on a 128GB DGX Spark.
76%
Reactions recovered by Terray's tuned model
Share of annotated reactions recovered on the RxnScribe diagram-reading test, per Terray and NVIDIA.
81%
Broken simulations KTH's agent repaired
Failed GROMACS molecular dynamics setups fixed in one overnight run on a single DGX Spark, each verified by GROMACS itself.
61%
Lila Sciences quantum-dot synthesis accuracy
Reported against 54% for GPT-6 Astra, at 10 to 20 times lower inference cost.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

People Involved

Organizations Involved

Timeline

October 2026 November 2026

3 events Latest: Today
  1. NVIDIA launches Nemotron for Science

    Today Release

    NVIDIA packages the model with LitmusBench v0.2 and post-training recipes, aimed at labs building their own scientific agents.

  2. Partners report benchmark gains from tuning

    Today Result

    Terray, Lila Sciences and Therna describe chemistry, lab-protocol and mRNA results; all figures come from the companies and NVIDIA.

  3. Nemotron 3.5 Super VL lands on Hugging Face

    Release

    NVIDIA posts the 120-billion-parameter open vision-language model under the OpenMDW-1.1 licence, with the chemistry dataset record following a day later.

Scenarios

1

Outside lab reproduces a Nemotron chemistry win

Possible Resolves by Q2 2027

Discussed by: The Frontier Wire and HPCwire both note the results are partner- and vendor-reported; NVIDIA has not published an independent evaluation.

A university or neutral benchmarking group tunes Nemotron 3.5 Super VL on LitmusBench or its own chemistry data and publishes a head-to-head against a frontier model. The trigger is the dataset being freely downloadable and small enough to fine-tune on modest hardware. If the gains hold, the ownership argument moves from marketing to measurable.

2

A partner result clears peer review

Possible Resolves by End of 2027

Discussed by: Trade coverage has treated the Terray, Lila and Therna numbers as company claims pending publication.

One of the three partner companies submits its tuned-model result to a journal or a conference with a reviewing committee, and the claim survives scrutiny. Peer review would also force disclosure the press release skipped, such as test-set sizes and baseline configurations.

3

NVIDIA extends the package to a second science

Likely Resolves by End of 2027

Discussed by: NVIDIA's own framing describes Nemotron for Science as a template of model, dataset and recipes rather than a one-off chemistry release.

NVIDIA ships a Nemotron for Science bundle for another domain, such as protein design or materials, with its own open dataset and tuning recipes. That would show the chemistry release was a repeatable product pattern rather than a single benchmark play.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

November 2020

AlphaFold at CASP14 (2020)

DeepMind entered a model in a biennial protein-structure competition and predicted shapes with accuracy researchers had not expected for years. The team then released code in 2021 and a public database of predicted structures, turning structure prediction from a specialist project into a tool anyone could query.

Then

Structural biologists gained a fast first guess for proteins that had never been crystallized.

Now

Prediction became infrastructure: a database anyone could search, built on one lab's model.

Why this matters now

AlphaFold established the pattern NVIDIA is copying: one strong model, a shared benchmark, public weights, and an argument that access beats secrecy.

February to July 2023

Meta releases LLaMA weights (2023)

Meta gave researchers access to a family of language model weights, which spread widely within days, then followed with a version carrying a commercial licence. Teams with modest budgets could fine-tune a capable model on their own data instead of paying per query.

Then

A wave of tuned variants appeared, most trained by groups with no access to frontier compute.

Now

Open-weight models became a standing alternative to hosted APIs, and a licensing question rather than a technical one.

Why this matters now

NVIDIA's ownership pitch, and the partner claims of matching a frontier model at a fraction of the cost, are the same argument applied to chemistry.

2009 to 2012

ImageNet and the shared benchmark era (2009-2012)

Researchers assembled a large labeled image dataset and ran a public competition on it. Because everyone measured against the same yardstick, progress became visible and comparable, and a deep network cut the error rate sharply in 2012.

Then

One benchmark made results from different teams directly comparable for the first time.

Now

The dataset and leaderboard set the direction of computer vision for a decade.

Why this matters now

LitmusBench v0.2 is an attempt to build that shared yardstick for chemistry reasoning. Whether others use it decides whether it works.

Sources

(7)