A title card for NVIDIA Nemotron for Science, headlined 'what shipped, and what it proves', with four statistics: roughly 50,000 LitmusBench v0.2 chemistry questions, 120B total and 12B active parameters, 81% of GROMACS setups repaired by a KTH agent in one overnight DGX Spark run, and a 1M-token context window.
Engineering & Research

Inside NVIDIA Nemotron for Science: What SC26's Open Chemistry Stack Actually Ships

Author

Gideon Frost

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Eighty-one percent is the number worth starting with, and it is not a benchmark score. It is the share of deliberately broken GROMACS molecular-dynamics setups that an agent built on NVIDIA Nemotron for Science repaired in a single overnight run on one DGX Spark workstation, per KTH Royal Institute of Technology — with every fix checked by GROMACS itself rather than by a model grading its own homework. NVIDIA announced the collection on 16 November 2026 at SC26 in Chicago, through Ian Buck, its VP of Hyperscale and High-Performance Computing. It is three artefacts in one box: the open Nemotron 3.5 Super VL checkpoint, an open chemistry dataset called LitmusBench v0.2 at roughly 50,000 questions, and a set of post-training recipes that wire the model into CUDA-X accelerated scientific tooling.

That combination is the interesting part, and it is not the part the headline sells. A vendor shipping another open checkpoint is routine by late 2026; NVIDIA shipped this one in October. What is new on 16 November is the scaffolding around it — the evaluation set, the recipes, and a set of named institutional deployments that describe what happens when you point the model at a real scientific workflow instead of a leaderboard. This piece works through what is actually in the collection, what the partner numbers do and do not establish, and where the gaps are for anyone planning to build on it.

The three artefacts, separated

NVIDIA's press release and its accompanying X post describe a single package. Unbundling it matters, because each piece has a different licence posture, a different maturity level, and a different reason to care.

• The checkpoint — Nemotron 3.5 Super VL, a 120-billion-parameter mixture of experts with 12 billion active per token, a Mamba-2 and Transformer hybrid with 512 routed experts per MoE layer, a C-RADIO vision encoder, and up to 1M tokens of context. Text, images and 2-minute MP4 video in; text out. OpenMDW-1.1, commercial and non-commercial use, weights on Hugging Face. This is not new on 16 November — the repository went up on 12 October and NVIDIA's own announcement followed on 15 October.

• The dataset — LitmusBench v0.2, an open set of roughly 50,000 chemistry-reasoning questions drawn from molecular property data and experimental drug-discovery records. It is built for the two jobs a scientific benchmark usually has to split between two files: it evaluates chemistry reasoning, and it is usable as post-training data. Solubility prediction and metabolic fate are the two task families NVIDIA names explicitly.

• The recipes and tooling — post-training recipes for domain adaptation, plus the wiring that lets an agent call CUDA-X accelerated scientific software as tools. This is the piece that turns a chat model into something that can run a simulation, read the output, and decide what to do next.

The checkpoint is the least novel of the three and the dataset is the most durable. Benchmark suites age; a 50,000-question chemistry set with a permissive licence is the sort of thing that is still being cited in three years, and it is the part of this release NVIDIA spent the most to give away.

A two-column summary card contrasting what was announced at SC26 with what is not yet established: the checkpoint is Nemotron 3.5 Super VL at 120B total and 12B active, the dataset is LitmusBench v0.2 at about 50,000 questions, partner results include 76% RxnScribe recovery at Terray and 61% versus 54% quantum-dot accuracy at Lila, the licence is OpenMDW-1.1, and no NVIDIA rate card or agent completion rate has been published.

What the partner results actually demonstrate

Seven organisations are named in the announcement, and the figures attached to them are the closest thing in the release to evidence. All of it is vendor-reported and none of it has been independently reproduced — that caveat applies to every number in this section, and it is the reason to read them as case studies rather than as scores.

Terray Therapeutics post-trained the model on chemistry diagrams — the reaction schemes drawn in patents and papers — to extract reactants, products and reaction conditions. On RxnScribe it recovered about 76% of annotated reactions and beat prior published results on two benchmarks. Diagram extraction is unglamorous and it is where a large share of real chemistry labour actually goes, so a 76% recovery rate on a named benchmark is a more useful claim than a reasoning score.

Lila Sciences is the number the release leans on. Generating execution-ready lab protocols for automated labs, the base model matched frontier performance with no fine-tuning at all. With lab-feedback post-training, quantum-dot synthesis accuracy reached 61% against 54% for GPT-6 Astra, at 10–20× lower inference cost. Two things are worth flagging. The seven-point gap is real if the protocol is, and the cost multiple is the more consequential figure — a 10–20× inference saving changes which experiments are worth running at all. But this is a comparison run by the party that did the post-training, on a task they chose, against a competitor they named.

Therna post-trained for agentic mRNA therapeutic design and nearly doubled protein-expression-range prediction accuracy. In a blinded comparison, scientists preferred its explanations over the base model's in 96% of cases. Preference is not accuracy, and a 96% preference rate is a statement about how the reasoning reads, not about whether it is right.

Then the institutional deployments, which are the part I find most convincing precisely because they are not framed as wins. Argonne National Laboratory is folding Nemotron into StructBioReasoner, a multi-agent system for designing protein-based cancer therapeutics. Jülich Supercomputing Centre is integrating it into drug and science discovery agents. RIKEN is serving the models, Super VL included, on its RIKYU supercomputing platform through the RiVault inference gateway — a real production path, on real hardware, for real users. And KTH built the GROMACS repair agent behind that 81% figure, on a single DGX Spark, overnight.

The KTH result deserves a sentence of its own because of its shape. A broken molecular-dynamics setup is a mundane, high-volume failure: a mismatched force field, a topology that does not match the coordinate file, a missing include. Fixing 81% of them is not a reasoning benchmark. It is the model doing the kind of work a computational chemistry group currently assigns to a graduate student, and it is verified by the simulation engine rather than by the model, which makes it the cleanest evidence in the release.

Why this is a post-training story wearing a model's clothes

Read the release again and notice what NVIDIA is actually asking you to do. Every single partner result involves adaptation: Terray post-trained on diagrams, Lila post-trained on lab feedback, Therna post-trained on mRNA design. Not one of the headline numbers comes from the checkpoint as shipped. That is deliberate, and it is the most honest thing about the announcement.

The argument is that base models are now good enough that the differentiator is the post-training loop, not the pretraining run. NVIDIA is selling the loop: a permissively licensed 120B checkpoint you can fine-tune without a legal review, a chemistry dataset to fine-tune on, and recipes that describe how to do it. That is a genuinely different pitch from "our model is smarter," and it is a pitch that only works if the weights are open — which is precisely why the licence matters more than the benchmark table.

It also explains the hardware line. NVIDIA says the model is tuned to run locally on a 128GB DGX Spark, and the BF16 build wants 4× H100 or 2× GB200 while the NVFP4 build fits a single DGX Spark, an H200, two H100s or a GB200. A post-training loop is an iterative, high-volume, low-latency workload, and the difference between "rent a node" and "there is a box on the bench" is the difference between a loop you run twice a quarter and one you run twice a day.

The gaps, stated plainly

Three things are missing from the release, and each would change how you plan.

• No independent evaluation of the science claims. Artificial Analysis does track Nemotron 3.5 Super VL — it lists an Intelligence Index of 14, output at 291 tokens per second, an 11.85-second time to first answer token, and a listed price of $0.20 per million input and $0.50 per million output with a 75% cache discount. That is useful and it is third-party. None of it is a chemistry benchmark. Every science number in the SC26 release comes from a party with a stake in the result, and LitmusBench v0.2 is new enough that no external group has published a score on it yet.

• No published price for the model as a service. The AA-listed $0.20/$0.50 is a third-party platform's meter, not an NVIDIA rate card — NVIDIA has published no per-token figure for NIM. If you are budgeting, the honest answer today is that self-hosting is the only path with a known cost.

• No evaluation of the agent loop, only of the components. The recipes describe an agent that plans, runs simulations, checks results and adjusts. Nobody in the release reports how often that loop completes without a human, how many tool calls it burns, or what it costs per completed task. Those are the numbers that decide whether an agent is a product or a demo, and they are absent.

A screenshot of NVIDIA's newsroom press release announcing Nemotron for Science at SC26, showing the announcement headline, the dateline and the opening paragraphs describing the collection of open models, datasets, tools and post-training recipes.

What this means if you are building

The practical read is that Nemotron for Science is not a model you adopt; it is a starting kit. If your group does chemistry or structural biology and you have GPU capacity, the combination of a permissively licensed 120B checkpoint, 50,000 chemistry questions and a set of published recipes is a genuinely good week-one package — and the KTH result suggests the tool-calling loop works well enough to be useful on mundane failures.

What it is not is a drop-in replacement for a hosted API. Nemotron 3.5 Super VL is not on our catalogue; we do not serve it today, and the self-hosted path is the one this release is actually built around. Where teams run a hosted model as the reasoning layer of a larger pipeline — and most do, because the agent loop around the science model is usually a general model — the routing question is separate from this announcement. One key across more than 200 models at provider list price with 0% markup, and automatic failover when an upstream degrades mid-run, is the part that keeps a multi-hour agentic job from dying at hour three because one provider had a bad afternoon.

What to watch next

A first external LitmusBench v0.2 score settles more than any single benchmark would, because it is the only number here that a neutral party can generate. A published NIM rate card would make the cost comparison arithmetic instead of positioning. And a report on the agent loop itself — completion rate, tool-call budget, cost per finished task — is the thing that would turn this from a promising kit into a procurement decision.

Until then, the fairest summary is the one the release accidentally writes for itself: NVIDIA shipped a good checkpoint in October, and in November it shipped the thing that makes the checkpoint useful. That is a less exciting announcement and a more valuable one.

A three-row card listing what the release does not establish: no independent science score despite Artificial Analysis tracking Nemotron 3.5 Super VL at Intelligence Index 14 and 291 tokens per second; no NVIDIA per-token rate card, only a third-party meter at $0.20 and $0.50 per million tokens; and no agent-loop completion rate, tool-call budget or cost per completed task.