# Nvidia Nemotron 3.5 Super VL explained: what you need to know

> Nvidia released open model Nemotron 3.5 Super VL and claims tuned versions beat rivals on specialised tasks.

*Nvidia has released an open 120B-parameter model that reads documents and images, and says customised versions beat Claude Opus 5 and GPT-6.1 Sol on specialised tasks.*

By Bez · SuggestedTech
Canonical: https://suggestedtech.com/news/nvidia-nemotron-3-5-super-vl-explained-what-you-need-to-know

Nvidia has released Nemotron 3.5 Super VL, an open model that can read documents and images. The company says versions of it, customised for specific jobs, beat Claude Opus 5 and GPT-6.1 Sol on several specialised tests. Those claims have not been independently checked.

## What happened

Super VL is a 120B-parameter model. "Parameters" are the internal settings a model learns during training, and more of them generally means a bigger model. It is a vision-language model, meaning it handles pictures as well as text. Nvidia built it for PDFs, charts, tables, images and webpages, and it can follow instructions and use tools.

Nvidia's headline claims, [reported by VentureBeat](https://venturebeat.com/technology/nvidias-new-open-nemotron-3-5-super-vl-with-post-training-beats-claude-opus-5-gpt-6-1-sol-on-specialized-task-benchmarks), cover five tasks. They apply to versions that went through post-training, the stage where a model is tuned for a particular purpose. Two examples:

- Medical claims accuracy rose from about 28% to about 72%, against about 38% for Claude Opus 5.
- Security attack resilience rose from about 35% to about 76%, against about 66% for GPT-6.1 Sol.

Gains across the five tasks ranged from 9% to 44%. This is not a claim that the standard Super VL beats those models in general. Nvidia has also not released task definitions, sample sizes or testing protocols, so the results can't be verified yet.

On speed, Nvidia says the model is up to 5x faster than a comparable open model, reaching about 625 tokens per second on four B200 chips. A token is a small chunk of text. On the Artificial Analysis index, an independent ranking, Super VL scores about 26. That trails GLM-5.3-Flash (42) and Qwen3.8-Flash-Next (40).

Nvidia says it trained the model on about 26 trillion tokens, then used 3.8 million post-training samples and reinforcement learning (learning from feedback) across about 100 environments. It is also releasing its training recipes and datasets.

Alongside it, Nvidia released Bolt 74B-A9B, built for local coding and research agents. It has 74B parameters in total, with 9B active at a time. A smaller 4GB version, Bolt E4B, is due on 23 October.

## What it means for you

If you mainly use chatbots, nothing changes directly. This is a model for developers and companies to build on.

If you build with AI, the licence is OpenMDW, which allows commercial use, fine-tuning and redistribution. That means you can adapt the model to your own task, which is where Nvidia's strongest results come from.

The benchmark wins are Nvidia's own figures for customised versions. On the independent index, the standard model sits behind two rivals. Treat the claims as a sign of what tuning can do, not proof that Super VL is better overall.

Running it in production has a cost. Nvidia requires NVIDIA AI Enterprise, priced at $4,500 per GPU per year. That covers software only.

## How to try it

You can get the weights on Hugging Face. The model is also available through OpenRouter, ModelScope and NVIDIA NIM. For development, Nvidia offers a free hosted API with rate limits.

Context: Nvidia's pending acquisition of Hugging Face, valued at about $12.93 billion, is expected to close in the first half of 2027.

## Key takeaways

- Nvidia's new open model reads PDFs, charts, tables, images and webpages.
- The benchmark wins are for custom-trained versions, not the standard model.
- Free to download, but production use costs $4,500 per GPU per year.

## Sources

- [Nvidia's new open Nemotron 3.5 Super VL beats Claude Opus 5, GPT-6.1 Sol on specialized tasks](https://venturebeat.com/technology/nvidias-new-open-nemotron-3-5-super-vl-with-post-training-beats-claude-opus-5-gpt-6-1-sol-on-specialized-task-benchmarks) — VentureBeat, 2026-10-15