Nemotron 3.5 Super VL: Nvidia's 120 billion parameter open model, fast but with unverified accuracy claims

Nvidia released Nemotron 3.5 Super VL on October 15, 2026, a 120 billion parameter open model that activates 12 billion parameters per token and handles a 1 million token context. Nvidia's own charts say post-trained versions beat Claude Opus 5 and GPT-6.1 Sol on narrow tasks, but the company has not published the test details.

The speed numbers are the firmer part of the launch. The accuracy comparisons are the part to read with a pencil, and VentureBeat's report, our main source for the benchmark detail, says as much.

What Nvidia shipped

The family has three members. Nemotron 3.5 Super VL is the 120 billion parameter model for documents, charts, tables, images and webpages, and it can follow instructions and use tools. Nemotron 3.5 Bolt 74B-A9B has 74 billion total and 9 billion active parameters, with native 2-bit compression that lets it fit in 24 GB of memory. A smaller Bolt E4B fits in 4 GB, targets voice assistants and on-device agents, and is due October 23.

Per Nvidia, training used about 26 trillion tokens across pretraining and continued pretraining, 3.8 million post-training samples and reinforcement learning across roughly 100 open environments. Nvidia's Megatron Bridge documentation describes Super VL as a hybrid language decoder paired with a RADIO vision encoder, a vision projector and a separate temporal video embedder.

The license and the cost of running it

VentureBeat reports an OpenMDW license that permits commercial use, fine-tuning and redistribution. We could not confirm that license text independently for Super VL, so treat it as the outlet's report. Weights are on Hugging Face in an NVFP4 variant, along with OpenRouter, ModelScope and Nvidia NIM, and Nvidia offers a free, rate-limited hosted API.

Production use is where the price appears. Nvidia quotes a $4,500 per GPU per year NVIDIA AI Enterprise license for NIM microservices, which excludes hardware, hosting and operations. "Open" here does not mean free to run at scale.

The speed claim is the solid one

Nvidia says Artificial Analysis measured about 625 output tokens per second for Super VL on four Blackwell B200 GPUs with 100,000 input tokens, at an Intelligence Index score of about 26. A comparison model, Inkling Small, ran at about 125 tokens per second at a similar score. Other open models in the index include GLM-5.3-Flash at 42 and Qwen3.8-Flash-Next at 40.

Two cautions. The index combines 10 evaluations and does not score image inputs separately, so a vision-language model's index score understates what it was built for. And the index scores of 40 or higher for rivals mean Super VL is faster, not smarter, than the models near it. That fits what VentureBeat concluded about the speed chart: it "establishes a generation-speed comparison under its stated test conditions."

For how to read a leaderboard score like that, see our explainer on what AI leaderboards rank and what a top score hides.

The accuracy chart is Nvidia's alone

The headline claim comes from a five-task chart covering document intelligence, security attack resilience, financial receipt extraction, scientific experimentation and medical insurance claims. Post-training gains range from 9 percent to 44 percent. In two tasks the report gives figures: on medical insurance claims, the base Super VL scored roughly 28 percent and the post-trained version roughly 72 percent, against about 38 percent for Claude Opus 5. On security attack resilience, roughly 35 percent rose to roughly 76 percent, against about 66 percent for GPT-6.1 Sol.

Then the caveats. VentureBeat's Carl Franzen wrote that "NVIDIA has not supplied the task definitions, sample sizes or testing protocol with the chart," and that "These are different customized versions on different specialized tasks." The article does not say who ran the evaluations.

That matters because the comparison is between a model fine-tuned for one task and general-purpose models that were not. A specialist beating a generalist on the specialist's home turf is expected and says little about the base model.

Why Nvidia is doing this

Nvidia sells chips, and open models that run best on its chips are good for chip sales. The model is optimized for Blackwell and Hopper, and can also run on Ampere GPUs. Named partners are customizing it, including Scale AI for Skydance and EY, Mercor for coding and Ramp, and others are integrating it, among them Atlassian, CrowdStrike, Salesforce and Synopsys.

The training recipes and datasets are released, with more datasets planned. That is the most useful piece for outside researchers, and it is more than most open-weight releases include.

What to do with it

If you process documents and charts at volume, test Super VL on your own files against whichever model you use now, measuring accuracy and cost per page. If you are choosing a model off the vendor chart, don't.

What independent testing would settle

Three outside checks would turn the claims into facts. A full Artificial Analysis entry that covers the Bolt variants as well as Super VL would show how the small models compare. An independent document-understanding test on real invoices, charts and scans would show whether the post-training gains carry over. And a published attack-resilience protocol would let security teams see what "security attack resilience" measured. Nvidia has released training recipes and datasets, which makes outside replication possible, and the labs and startups customizing the model are the likeliest to publish results.

Our take

Nemotron 3.5 Super VL looks like a fast, usable open model with unusually good release materials. The "beats Claude Opus 5 and GPT-6.1 Sol" line is not established by anything outside Nvidia, and Nvidia itself has not shown the test. We would take the speed number seriously, run our own accuracy test and ignore the specialist-versus-generalist headline until the protocol is published.