Article / Public
NVIDIA announces Nemotron 3.5 Super VL and Bolt models
NVIDIA introduced the 120-billion-parameter Super VL and locally focused Bolt models. The announcement emphasizes visual document processing and different memory targets; performance statements remain producer claims.
NVIDIA announced the release of Nemotron 3.5 Super VL, an open 120-billion-parameter model, on October 15, 2026. The company is also adding Bolt models to the family for local agent workloads. The announcement focuses on enterprise agents working with visual documents and local applications designed for more limited memory.
According to NVIDIA, Super VL and Bolt 74B-A9B 24GB are available now. Bolt E4B 4GB is scheduled for October 23, 2026. That timetable does not mean all three models are already accessible. The information below comes from NVIDIA’s announcement, not an independent product test.
Super VL focuses on information in visual documents
NVIDIA describes Super VL’s vision-language capabilities in terms of optical character recognition, tables, charts and structured information in visual documents. In the company’s account, the model extracts text from web pages, reads graphs and charts, and produces insights from content found in documents such as PDFs.
The underlying approach combines text and visual content in one model. NVIDIA says this can reduce the need to route agents between separate systems as they move from reading documents to subsequent actions. However, that description does not establish equal accuracy across every document layout or task. The supplied evidence contains no independent error rates broken down by document type.
The 120-billion-parameter figure describes model scale; it does not independently establish success on a particular task. Evaluating the announcement therefore requires treating model size and the document-understanding capabilities an application actually needs as separate questions.
From visual documents to information
- Document input
Web pages, PDFs and visual documents are among the model’s described inputs.
- Text and structure
Focus areas include optical character recognition, tables, charts and structured information.
- Using information
NVIDIA describes producing insights from retrieved content and using it in agent workflows.
Bolt targets local operation at different memory levels
A key technical detail for Bolt 74B-A9B 24GB is native 2-bit quantization. NVIDIA says the model is designed to fit in 24GB of memory using that method and run alongside other applications. The company positions it for everyday local agent tasks on NVIDIA RTX and DGX Spark.
This should be read as a design target. The announcement does not provide independent measurements verifying available memory or concurrent performance for every combination of applications. The 24GB figure is therefore not a guarantee that every workload will run smoothly on a particular computer.
NVIDIA describes Bolt E4B 4GB as intended for physical and conversational agents. Local voice assistants, interactive characters and embedded interfaces are among its examples. Unlike the two models already available, its announced availability date is October 23.
What does openness include?
NVIDIA says Super VL ships with open weights, open datasets and training recipes. The company also says organizations can post-train the model on their own data and deploy it on their own infrastructure. Listed customization tools include NVIDIA NeMo and Python-based pipelines.
These components suggest an approach different from simply accessing a ready-made service: organizations can have options to adapt the model and choose where it runs. However, “open” in this article describes the components NVIDIA says it provides. The supplied evidence does not include a detailed review of all licensing conditions.
The model family’s stated use cases
- Super VL
An open 120-billion-parameter model focused on visual documents and enterprise agent workflows.
- Bolt 74B-A9B 24GB
Described as designed to fit in 24GB of memory through native 2-bit quantization.
- Bolt E4B 4GB
For physical and conversational agents; availability is scheduled for October 23, 2026.
How should performance claims be read?
NVIDIA uses strong comparative language about speed, accuracy and efficiency. Although the source names some rankings and evaluations, the evidence here does not establish that those results have been independently reproduced. These statements should therefore be treated as producer claims, not results verified by Urazo.
Document-processing accuracy, fitting into local memory and running alongside other applications are separate evaluation questions. A positive result in one does not automatically establish the others. Assessing suitability requires evaluating the intended workflow with its own documents and operating conditions.
Urazo interpretation
The announcement’s clearest feature is its use of different models for visual document processing and local agent operation, rather than presenting one model as the answer to every use case. Super VL’s stated focus is document content; the distinguishing framework for Bolt is local operation and memory targets.
That distinction can provide a useful starting point for selection, but it does not independently guarantee accuracy, speed or privacy. Open weights and local deployment options can offer control, while implementation still matters. This article is based on NVIDIA’s October 15 announcement; Urazo has not tested these models.
Sources
Image: AI-generated illustration. Explanatory animations are conceptual diagrams, not measurements or product demos.