A title card for the comparison, headlined 'Nemotron for Science vs Nemotron 3.5 Super', stating that these are not two models: the October checkpoint at 120B total and 12B active with 512 routed experts per MoE layer, weights on Hugging Face 12 October 2026 and announced 15 October, the November package on 16 November at SC26 in Chicago, and roughly 50,000 LitmusBench v0.2 questions as the only new model artefact.
Engineering & Research

Nemotron for Science vs Nemotron 3.5 Super: Same Weights, Different Product

Author

jinhao.song

Date Published

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

This comparison has an awkward fact at its centre: NVIDIA Nemotron for Science and NVIDIA Nemotron 3.5 Super are not two models. They are the same 120-billion-parameter checkpoint, shipped five weeks apart, wearing different packaging. Nemotron 3.5 Super is the weights — the repository under the fuller name Nemotron 3.5 Super VL 120B A12B, up on Hugging Face since 12 October 2026, announced on 15 October, OpenMDW-1.1, 12 billion active parameters per token, 1M-token context. Nemotron for Science is what NVIDIA announced at SC26 on 16 November: that checkpoint plus LitmusBench v0.2, roughly 50,000 chemistry-reasoning questions, plus post-training recipes and CUDA-X tool wiring, plus seven named institutional deployments.

So the useful question is not which one is better. It is what the November package adds to the October checkpoint, whether the additions justify a separate name, and what you would actually download today if you wanted to run chemistry agents on NVIDIA silicon. That is a more interesting question than a spec table, and the answer is more flattering to NVIDIA than a repackaging story usually is.

What is identical, and it is most of it

Start with the checkpoint, because everything in both packages rests on it. The architecture does not change between October and November: a Mamba-2 and Transformer hybrid mixture of experts, 512 routed experts per MoE layer, multi-token prediction heads, a C-RADIO vision encoder and a vision-language connector, 120B total parameters with 12B active per token, and up to 1M tokens of context. Inputs are text, RGB images and MP4 video up to two minutes; output is text. Reasoning effort is configurable through the chat template at off, medium or on.

• Parameters — 120B total, 12B active. Identical in both. There is no Nemotron for Science checkpoint with a different parameter count, because there is no separate Nemotron for Science checkpoint.

• Licence — OpenMDW-1.1 for commercial and non-commercial use. Identical, and it is what makes the post-training pitch possible at all.

• Context — 1M tokens, with the card recommending video at 2 FPS up to 256 frames, or 1 FPS and 128 frames at 1080p.

• Hardware floor — 4× H100 or 2× GB200 at BF16; the NVFP4 build fits a single DGX Spark, one H200, two H100s or one GB200. The DGX Spark line is the one SC26 leans on, and it is the same line the October card already carried.

• Benchmarks — GPQA-Diamond 89.52, SWE-bench Verified 71.10, SWE-bench Pro 51.78, Terminal-Bench 2.1 60.67, tau2 retail 69.24, BrowseComp 78, MathVision 78.16, CharXiv reasoning 80.82. All vendor-reported, all published with the October card, none of them new.

If you already pulled Nemotron 3.5 Super in October, you have the model that Nemotron for Science is built on. Nothing in the November announcement changes its weights.

What is actually new, and it is not nothing

Now the part worth the article. Three things arrived on 16 November that did not exist on 15 October, and each changes what you can do on day one.

LitmusBench v0.2 is the headline. Roughly 50,000 chemistry-reasoning questions, built from molecular property data and experimental drug-discovery records, covering tasks like predicting solubility and how the body metabolises a compound. Two properties make it more than a leaderboard: it is open, and it is designed to be used for post-training as well as evaluation. A 120B model with a permissive licence and no chemistry data to fine-tune on is a model you have to build a dataset for; a 120B model that ships with 50,000 annotated chemistry questions is a model you can adapt in a week. That is the difference the package buys.

The recipes are the second piece. NVIDIA published post-training recipes for domain adaptation and documentation for embedding the model in agent workflows that read scientific information, write code and invoke CUDA-X accelerated tooling. For anyone who has tried to get a general model to call a molecular dynamics package correctly, the tool-wiring documentation is the unglamorous part that determines whether the loop works at all.

The deployments are the third, and they are evidence rather than artefacts. Terray Therapeutics post-trained the model on chemistry diagrams and recovered about 76% of annotated reactions on RxnScribe, beating prior published results on two benchmarks. Lila Sciences hit 61% quantum-dot synthesis accuracy against 54% for GPT-6 Astra at 10–20× lower inference cost. Therna nearly doubled protein-expression-range prediction accuracy for mRNA design, with scientists preferring its explanations in 96% of blinded comparisons. Argonne is integrating it into StructBioReasoner; Jülich into drug and science discovery agents; RIKEN is serving it on RIKYU through the RiVault gateway. And KTH built a GROMACS repair agent that fixed 81% of deliberately broken setups in one overnight run on a single DGX Spark, verified by GROMACS itself.

A two-column scoreboard contrasting the October checkpoint with the November package across what it is, when it was announced, dataset, recipes, evidence and price: the checkpoint is the open 120B model alone with a vendor benchmark table and no rate card, while the November package adds LitmusBench v0.2 at about 50,000 questions, post-training recipes with CUDA-X tool wiring, and seven named partners and institutions.

Why the split into two names is defensible

The cynical read is that NVIDIA announced the same model twice. I do not think that holds up, and the reason is that the two announcements answer different questions.

October's question was "what is this checkpoint and can it be deployed?" The answer was a 120B vision-language MoE with a 1M-token window, a hardware floor spanning a DGX Spark to a GB200 node, and a speed claim — 5× faster than the next best in its class, per NVIDIA, though Artificial Analysis did not list the model at the time the claim was made. That is a model launch.

November's question is "what does a chemistry group do with it on Monday?" That answer is a dataset, a recipe, and seven organisations who have already done it and reported numbers. Shipping weights is not the same product as shipping a workflow, and the second one is what a research group can actually adopt.

The strongest evidence for that reading is the shape of the partner results. Every single headline number comes from a post-trained variant, not from the checkpoint as shipped. Terray adapted it on diagrams. Lila adapted it on lab feedback. Therna adapted it on mRNA design data. The package is not claiming the base model is a chemist; it is claiming the base model is a good enough starting point that a domain team can make it one. That is a more modest claim and a more useful one, and it would be strange to make it in October, before the dataset existed to make it with.

What the package still does not give you

Three gaps, in order of how much they would cost you to discover the hard way.

• No independent chemistry evaluation. Artificial Analysis does list Nemotron 3.5 Super VL — Intelligence Index 14, 291 tokens per second output, 11.85s to first answer token, $0.20 per million input and $0.50 per million output with a 75% cache discount. That is a third-party general score, and it is worth knowing. It is not a chemistry score. Every science figure in the SC26 release was produced by a party with a stake in it, and LitmusBench v0.2 is new enough that no outside group has published a result on it.

• No vendor price. The $0.20/$0.50 above is a third-party platform's meter, not an NVIDIA rate card. NVIDIA has published no per-token NIM pricing, which means the only budgeting path with a known cost is self-hosting.

• No completion rate for the agent loop. The KTH result reports 81% of test cases fixed, which is the most concrete number in the release — but it does not report how many tool calls each fix took, how often the agent looped without progress, or what a completed task cost. Those are the numbers that determine whether an agent is deployable or merely impressive.

How to run it, and where the surrounding stack fits

The deployment advice is unusually clear for an open-weights release, so take it literally. The card specifies vLLM with TP4 or TP2, --trust-remote-code, async scheduling, MTP speculative decoding with five draft tokens and --max-num-seqs 32; or SGLang with FP8 KV cache and EAGLE speculative decoding at five steps. Sampling is temperature 1.0 with top_p 0.95 for text and 0.6 with 0.95 for vision, and the server name in the examples is nemotron-3.5-super-vl. Multimodal limits are one image and one video per prompt. That is a configuration you can copy rather than derive.

Nemotron 3.5 Super VL is not on our catalogue — we do not serve it, so the self-hosted path is the real one, and that is a fair thing for an article about an open checkpoint to say. Where our side of the stack does come in is around the science model rather than in it. A research agent is a planner wrapped around a domain checkpoint: it decomposes a task, calls the chemistry model, calls a simulation, reads the result, and decides what to do next. The planner is usually a general model, and that is the layer where one key across more than 200 models at provider list price with 0% markup saves the integration work — with automatic failover, so a six-hour overnight run does not fail at hour five because one provider degraded. Model fusion is the other lever worth knowing about here: a panel of models answering the same chemistry question together is a reasonable way to triage which protocols are worth running on real hardware.

A five-row deployment card titled 'How to run Nemotron 3.5 Super VL', giving the vLLM configuration with TP4 or TP2 and MTP speculative decoding at five draft tokens, the SGLang configuration with FP8 KV cache and EAGLE at five steps, sampling temperatures of 1.0 and 0.6 with top_p 0.95, the one-image and one-video per-prompt limit, and the hardware floor of 4x H100 or 2x GB200 at BF16 or a single DGX Spark on the NVFP4 build.

What to watch

A first external LitmusBench v0.2 score is the single most useful thing that could appear, because it is the only number in this story a neutral party can generate. A NIM rate card would make the hosted-versus-self-hosted comparison arithmetic rather than guesswork. And an agent-loop report — completion rate, tool-call budget, cost per finished task — would convert the KTH and Lila results from encouraging anecdotes into something a group can plan against.

Until then, the short version: if you pulled Nemotron 3.5 Super in October, Nemotron for Science does not replace it, it extends it with the data and recipes that make it a chemistry tool. If you did not, the November package is the better entry point, because it is the same weights with a starting kit attached.

A screenshot of the Hugging Face model card for nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16, showing the repository name, the model card header and the opening sections of the documentation.