An agent built for messy information
In its report on the launch, VentureBeat described NVIDIA’s Nemotron 3.5 Super VL as an open model aimed at agents that can work across documents, charts, images, webpages and tools. The model is meant to do more than answer a question about a file. It can interpret visual information, reason over it, follow instructions and call external tools as part of a multistep workflow.
NVIDIA’s announcement says Nemotron 3.5 Super VL has 120 billion total parameters, with 12 billion active during operation. That distinction allows the model to offer the capacity of a much larger system while limiting the amount of computation used for each response.
The intended applications are practical and document-heavy. A financial institution could use the model to review a credit agreement. An insurer could process a claim by combining scanned records, photographs and policy language. A company could ask it to check internal rules, gather supporting evidence and produce a recommendation for a human reviewer.
That workflow reflects a shift in how companies are evaluating AI. The question is no longer simply whether a model can write a convincing answer. It is whether the model can behave like a junior analyst who works across disconnected sources, follows a company’s procedures and leaves enough of a trail for someone else to inspect the result.
Open weights, customized procedures
NVIDIA’s research page lists Super VL among the company’s Nemotron 3.5 releases and points to its technical documentation and evaluation methods. The broader lineup also includes Nemotron 3.5 Bolt 74B-A9B, positioned for local coding, research and multistep agent tasks. NVIDIA has said a smaller Bolt E4B model, with a footprint of about 4 GB, is scheduled for October 23 for conversational and device-based agents.
The strategic appeal is control. NVIDIA says organizations can post-train Super VL using their own documents, tools and policies, then deploy it on infrastructure chosen around security, data residency and operational constraints. For companies handling medical, financial or legal information, that can be more attractive than sending every document to an outside hosted service.
The model’s NVFP4 repository on Hugging Face provides a released checkpoint and reports results across document understanding, multimodal, coding and agentic evaluations. NVIDIA has also claimed that, after post-training, the model exceeds Claude Opus 5 and GPT-6.1 Sol on selected specialized benchmarks.
Those comparisons require care. A strong result on a narrow enterprise test can show that an adapted open model is useful for a particular workflow. It does not establish broad superiority in reasoning, reliability, speed or cost. In enterprise settings, a model that is slightly less capable but easier to keep inside a company’s security boundary may still be the more valuable system.
The uncomfortable evidence
The most important challenge to NVIDIA’s presentation comes from an evaluation that tested whether the model could notice problems rather than simply answer visible questions. Plotly reported that Super VL passed 30 of 43 questions on its private data analytics exam. But Plotly said the model missed all of the test’s planted data-quality issues and hidden-needle problems.
That finding goes to the heart of the enterprise promise. An agent reviewing a spreadsheet or claims file may be asked to calculate a result, but a dependable employee would also ask whether records are missing, whether a column is inconsistent or whether an important exception is buried elsewhere. An agent that confidently completes the assigned task while failing to search for such defects could make an organization faster at producing the wrong answer.
Plotly’s criticism is not proof that Super VL cannot support serious work. It is evidence that benchmark success may conceal a gap between responding to a defined prompt and independently investigating the quality of the information behind it. That distinction matters most in high-stakes operations, where the absence of a warning can be more dangerous than an obvious factual error.
CellCog’s comparison of Super VL and Bolt likewise presents the models in terms of practical trade-offs rather than a simple ranking. Larger systems may handle more demanding multimodal and agentic tasks, while smaller models can be easier to run locally and at lower resource requirements.
A test of trust
NVIDIA’s training recipe documentation describes the training and post-training process, while noting that the supervised fine-tuning and reinforcement learning data used for Super 3.5 has not been publicly released. That limits outsiders’ ability to reproduce every part of the company’s results.
Nemotron 3.5 therefore arrives with a compelling proposition and a clear burden of proof. Open models could give enterprises more freedom to adapt AI to internal rules without surrendering control of sensitive data. But the decisive question will be whether these agents can identify missing evidence, obey specialized policies and explain their decisions well enough for people to trust them.
For now, NVIDIA has supplied the machinery. The harder task is teaching an artificial analyst when not to proceed.
- Coolcaesar · CC BY-SA 4.0
Daniel Reyes is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Daniel Reyes's usual length and in Daniel Reyes's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.