NVIDIA Nemotron 3.5 Super VL: what the chart shows

On 15 October 2026 NVIDIA released Nemotron 3.5 Super VL, an open model that reads text and images. Social posts since then say it beats GPT-6 and Claude Opus 5 at specialized work once it is trained on your data. We traced that claim to a chart in NVIDIA's developer blog. The numbers in the posts match the chart. What the chart can and cannot support is narrower than the headline.

What NVIDIA confirms about Nemotron 3.5 Super VL

These points come from NVIDIA's announcement, its developer blog and the model card on Hugging Face, all dated 15 October 2026.

Two details to keep in mind. The 1M-token figure is the model's maximum. The card's TensorRT-LLM setup for a DGX Spark sets a maximum sequence length of 131,072 tokens, so a single small machine will not necessarily run the full length. And "open" does not mean "no terms": read the OpenMDW licence before building on it. Our piece on open-weight and paid models explains the difference.

What NVIDIA's chart shows

Figure 2 in the developer blog is titled "Post-Training Delivers Leading Accuracy for Specialized Agents". For five tasks it plots the base model, the same model after a partner trained it on its own data, and one named model from OpenAI or Anthropic. NVIDIA prints only the gain for each task. The bar heights below are our readings of the chart, rounded to the nearest point, so treat them as approximate.

How the headline claim holds up

The post says the model trailed the frontier models on all five tasks out of the box, then led all five after partners trained it. The chart agrees: the base bar sits below the named model on every task, and the trained bar sits above it. The Scale AI figures also match. The chart puts document work at about 86% against about 83% for Claude Opus 5, which is the 86% against 83% in the post.

Three things the headline leaves out:

• It is five tuned models, not one. Each result is a different version trained by a different partner for one task. VentureBeat, which covered the launch, put it this way: "These are different customized versions on different specialized tasks, not a claim that the base Super VL model beats those models generally."
• Each task has a different rival. Claude Opus 5 is the comparison on two tasks. On the others the chart names GPT-6.1 Sol, GPT-6 Luna and GPT-6 Astra. No task compares one model against all of them.
• The margins vary. The trained model is ahead by about 3 points on document work, about 3 on receipts, about 6 on science, about 10 on security and about 34 on medical claims. A gap of 3 points is small enough that test noise and our chart reading both matter.

The cost claim is separate from the chart. The developer blog says Scale AI post-trained the model for automated vendor credit processing at Skydance, and that inference cost for this workload "will decrease by 40x and latency by 5x". It is written as an expectation, and the blog does not say which model the 40x is measured against. The post places it beside the Claude Opus 5 comparison. NVIDIA's text does not.

What is still missing

The post ends: "NVIDIA hasn't shared the test details yet, so run your own evals before switching." That is accurate. VentureBeat reports that NVIDIA "has not supplied the task definitions, sample sizes or testing protocol with the chart". We found nothing more in the model card or in the technical report NVIDIA links. The report describes how the base model was built and tested. It has no partner results. Scale AI's blog does not list a write-up of this work either.

So the chart tells you what NVIDIA and its partners say they measured. It does not tell you how many examples there were, how the answers were scored, or whether the comparison models were prompted as carefully as the tuned ones.

What this means if you run AI work

A model that can be trained on your own data is a different purchase from a general model you rent. If a smaller open model, tuned on your documents, matches a larger paid one on your task, the cost per request can fall and you can choose where it runs. That is a reason to test it. It is not yet a reason to switch.

• Run your own evaluation. Use your documents and your own pass mark, on the base model, the tuned model and the model you use today.
• Read vendor and partner numbers as claims. Check what was measured, on what data and against which model version.
• Price the whole job. Training on your data, hosting and monitoring cost money and people. See what it takes to run an open model yourself.

We track how model prices and capabilities change in Model Watch.

Starfield is not affiliated with NVIDIA. This article is general information and not investment advice.