NVIDIA Nemotron 3.5 Super VL: Frontier level accuracy at one fifth the size
Follow Us
For many years in AI, the answer to everything has been bigger. Larger numbers of parameters, more GPUs, and larger bills. NVIDIA is saying the exact opposite. Its newest offering, the Nemotron 3.5 Super VL, has 120 billion parameters but claims it performs on par with its largest reasoning model on multimodal agentic tasks using about one-fifth the number of parameters. It also says it is five times faster than any other comparable open-source model on the market. Big talk. What does NVIDIA mean by frontier and who is using this model and what does it matter if you never set foot in a data center?
SurveyAlso read: Muse Spark Open: Open weights with a new model built for coding and AI agents
Frontier accuracy comes with an asterisk
This benchmark is against its own largest reasoning model, not all closed models out there. Take frontier as a code name and try it on your application. Nonetheless, those are precisely the skills that are being targeted. Super VL deals with instruction following, tool utilization, multi-step business workflow reasoning and visual reasoning. It is those down-to-earth skills that will determine whether an agent will be able to complete a task successfully or meander away. It was developed by NVIDIA in collaboration with the Nemotron Coalition. Referencing the Artificial Analysis Intelligence Index, it claims up to 5x the throughput compared to similar models. Speed is crucial. It is important to understand that agents call a lot of steps in a sequence, which means that any inefficiency will accumulate. Imagine an agent doing due diligence for a week or two.
It reads documents the way agents meet them
Also read: Amazon Bee AI wearable is here: What it does and why everyone is talking about it
The “VL” part of the name means “vision language,” and it is the secret star in this product. Super VL is designed for optical character recognition, tables, charts, and other structured information from visual documents. Give it a webpage or a PDF, and it will extract text and graph data and provide insights. Why is that important? Office workflows include lots of PDFs, scanned forms, and spreadsheets added as image snippets. Most existing agent architectures use one neural network for textual information and another for images and combine their outputs. NVIDIA envisions one neural network that will have context across document, data, and decision processes. Less context loss and integration steps mean less opportunities for errors. Some companies are already using the product.
Open weights put you in control
The operative word here is “open.” NVIDIA offers the weights, the datasets, and the training recipes for this model, which is pretrained using some 26 trillion tokens of training material and some 3.8 million post-training samples. The model is available on Hugging Face, OpenRouter and ModelScope, or can be deployed as a NIM microservice. If it’s an Indian bank or insurance company worried about their customer data going into an external API endpoint, that changes the discussion entirely. This is the distinction between leasing a model versus owning it. You could run it on your own infrastructure, tune it on your own datasets, and keep everything where your legal compliance people are comfortable. NVIDIA claims that the model is available across DGX Spark, DGX Station, RTX PRO, and Jetson as well as data centers, and that on Dynamo, it can handle 8x more users at 85% less latency. These numbers have to come with a grain of salt until there are independent evaluations. But you should be keeping an eye on the Bolt series as well, particularly E4B on October 23.
Also read: Maya 2 explained: A free open weight voice AI model for 22 Indian languages
