NVIDIA Launches Nemotron 3.5 Super VL, a 120B AI Agent Model

NVIDIA has released Nemotron 3.5 Super VL, a 120-billion-parameter open-weight artificial intelligence model designed to power enterprise agents that can understand documents, analyse images and video, reason through complex instructions and use external tools. Announced on October 15, 2026, the model combines several capabilities in one system, targeting businesses that want AI agents to handle extended workflows instead of responding only to individual prompts.

The model uses a mixture-of-experts architecture that activates approximately 12 billion parameters per token, despite having 120 billion parameters in total. It supports context lengths of up to one million tokens and accepts text, images and video as inputs, producing text responses. NVIDIA has also introduced smaller Nemotron 3.5 Bolt models for deployments closer to users, including local computing systems and edge devices. The release expands NVIDIA’s effort to provide customisable AI models and supporting tools across data centres, workstations and smaller hardware platforms.

Key Takeaways

New model: NVIDIA released Nemotron 3.5 Super VL on October 15, 2026.
Model size: It has 120 billion total parameters, with approximately 12 billion active per token.
Multimodal capabilities: The model processes text, images, documents and video, then generates text responses.
Long context: It supports context lengths of up to one million tokens.
Agentic capabilities: It is designed for reasoning, coding, instruction following and tool use across complex workflows.
Licence: The model weights use NVIDIA’s OpenMDW-1.1 licence, which developers should review before commercial deployment.
Additional releases: Nemotron 3.5 Bolt includes a 74B-A9B model for local agent workloads and an E4B model designed for lightweight conversational applications.

What Is NVIDIA Nemotron 3.5 Super VL?

Nemotron 3.5 Super VL is an open-weight vision-language model designed to do more than generate text from a prompt. It can interpret visual information, work with lengthy documents and use tools as part of a larger task.

Consider an enterprise workflow involving a financial report. A conventional text model might summarise the extracted text, while a separate vision model could be needed to interpret charts or scanned pages. An agent built around Nemotron 3.5 Super VL can combine document understanding, visual reasoning and tool-assisted operations within one model architecture.

The goal is to reduce the need to connect separate models for every stage of a task. A business could use the model to read a report, identify a discrepancy between a chart and a table, check relevant information through an approved tool and produce a summary for a human reviewer.

NVIDIA is targeting specialised enterprise agents that can remain active and perform multiple steps. Such agents may be useful in document processing, customer operations, research, software development and internal business workflows.

However, releasing a capable model does not automatically provide a complete enterprise agent. Developers still need to build the surrounding application, connect tools, set permissions, manage data and implement safeguards.

The 120B Architecture Activates Only 12B Parameters per Token

One of the model’s central technical features is its mixture-of-experts, or MoE, architecture.

Rather than activating every parameter for each token generated, an MoE model routes work through selected components called experts. Nemotron 3.5 Super VL has 120 billion parameters in total but activates approximately 12 billion per token.

Technical specification	Details
Total parameters	120 billion
Active parameters per token	Approximately 12 billion
Architecture	Hybrid Mamba-2, Transformer and mixture-of-experts
Context length	Up to 1 million tokens
Inputs	Text, images and video
Output	Text
Vision encoder	C-RADIO-based vision encoder
Additional efficiency feature	Native multi-token prediction

The distinction between total and active parameters matters because it affects computation during inference. Activating a subset of parameters can improve efficiency compared with a dense model of similar total size, depending on the architecture and workload.

It does not mean that the model requires only the memory needed for 12 billion parameters. The model’s full weights, visual components, runtime, context cache and other components still affect memory requirements. Running the full model locally can therefore require substantial specialised hardware.

NVIDIA’s hybrid design combines Mamba-2 sequence-processing layers with attention and expert-based computation. The objective is to balance model capacity, long-context processing and inference efficiency for demanding agent workloads.

Why a One-Million-Token Context Window Matters

A context window determines how much information a model can consider within a single interaction or processing sequence, subject to the system’s implementation.

Nemotron 3.5 Super VL supports up to one million tokens of context. This can be useful when a task involves long documents, large collections of records or multiple sources that need to be considered together.

For example, an enterprise agent might need to examine a long policy manual, compare it with several contracts and identify clauses that conflict with internal rules. A large context window can reduce the need to split the material into many separate requests.

It may also help developers build workflows in which the agent maintains more task information while using tools and processing results.

However, a large context window is not a guarantee of perfect recall or reasoning. Models can still overlook details, misinterpret relationships or produce incorrect conclusions when given large amounts of information. Long-context tasks also consume computing resources, and performance can depend on how the application organises the material.

Businesses will therefore need to test whether the model can reliably retrieve and connect relevant information from their own documents.

Multimodal AI Combines Documents, Images and Video

The VL designation highlights the model’s vision-language capabilities. Nemotron 3.5 Super VL can process text alongside visual inputs, allowing it to work with information that would be difficult to handle using text alone.

Potential applications include:

Document analysis: Extracting and interpreting information from reports, tables, forms and scanned documents.
Visual inspection: Analysing images to identify relevant objects, conditions or discrepancies.
Video understanding: Processing video frames to summarise events or examine a sequence of activities.
Research assistance: Combining information from written material and visual evidence.
Enterprise automation: Using interpreted information as input to a tool-assisted workflow.

The model produces text output rather than directly generating video or images. Its visual understanding is intended to support reasoning and decision-making within an application.

For instance, a business could build an agent that reviews a set of documents and associated images, flags inconsistencies and prepares a report for a human operator. A developer could also connect the model to approved databases or software tools so that it can retrieve information or perform permitted actions.

Such uses require careful evaluation. Visual models can misread charts, overlook small details or infer information that is not present. In consequential settings, businesses should use validation steps and human review rather than treating model outputs as automatically correct.

Nemotron 3.5 Super VL Is Built for AI Agents

The release is particularly focused on agentic AI, a category of systems that can break down tasks, use tools and carry out multiple steps toward a goal.

A conventional chatbot generally responds to a user’s request. An agent can follow a longer sequence: gather information, inspect the results, decide what to do next, call a tool and then use the returned information to continue.

Nemotron 3.5 Super VL is designed to support this kind of workflow through its reasoning, coding, visual understanding and tool-use capabilities.

A customer-support agent, for example, could inspect a submitted document, retrieve relevant information from an authorised system, compare the information with company policy and draft a response. A software-development agent could inspect screenshots or documentation, write code and use connected development tools to help complete a task.

The model does not independently gain access to every external service. Developers must supply the tools, define the available actions and configure access permissions. This is especially important when an agent handles customer records, financial information or other sensitive data.

Always-on agents also create operational challenges. Businesses need to monitor repeated tool calls, control costs, log actions and prevent agents from making unauthorised changes. The model is one component of the overall system rather than a complete substitute for application-level security.

NVIDIA Also Introduces Nemotron 3.5 Bolt

Alongside Super VL, NVIDIA announced two smaller models intended to extend agent capabilities beyond large data-centre deployments.

Model	Intended role	Availability announced
Nemotron 3.5 Super VL 120B-A12B	Multimodal enterprise agents	Available
Nemotron 3.5 Bolt 74B-A9B 24GB	Local agent workloads	Available
Nemotron 3.5 Bolt E4B 4GB	Lightweight conversational agents	Expected October 23, 2026

The Bolt models target different deployment constraints. The 74B-A9B version is designed for local agent tasks, while E4B targets compact systems and faster conversational applications.

Smaller models can be attractive where latency, hardware cost, privacy or local processing is important. But the best model depends on the task, available memory and required quality. A compact model may be more practical for a narrow workflow, while a larger multimodal model may be better suited to complex documents and visual reasoning.

The different model sizes also reflect a wider industry trend: AI deployment is moving beyond a single large model in a cloud data centre toward a mix of centralised and local systems.

What Developers Need to Know About Access and Licensing

NVIDIA has published model resources through its developer and model-hosting channels, including model cards and supporting materials. Developers can examine the model’s specifications and deployment requirements before deciding whether it fits their applications.

The weights are distributed under the OpenMDW-1.1 licence. Open-weight access allows developers to download and customise model weights under the applicable terms, but it should not automatically be equated with an unrestricted open-source licence.

Teams considering commercial use should review the licence, usage conditions and any relevant restrictions. They should also examine the hardware requirements, inference framework, data-handling requirements and operational costs.

NVIDIA has published training recipes and related tooling for the Nemotron family, but its documentation notes that some supervised fine-tuning and reinforcement-learning datasets used in the Super VL process have not yet been released. This distinction matters for teams that want to reproduce the complete training process rather than simply download and run the model.

Access to weights, training code, datasets and full reproducibility are separate questions. Developers should verify which components are actually available before planning a research or commercial project around them.

How Nemotron 3.5 Super VL Fits NVIDIA’s AI Strategy

NVIDIA is best known for its GPUs and data-centre platforms, but its software and model ecosystem is also central to its AI strategy. Open-weight models can encourage developers to build applications on NVIDIA-compatible infrastructure, creating demand for the hardware and software used to train and serve those models.

Nemotron 3.5 Super VL supports that approach by giving developers a model they can customise for specialised workflows. The smaller Bolt models extend the same family toward local systems and edge deployments.

The strategy also reflects the growing importance of agentic AI. Businesses are exploring systems that can perform multiple steps across internal applications rather than simply generate text. Models that can understand documents, reason and use tools are therefore competing on their ability to complete real workflows, not only on conventional chatbot benchmarks.

NVIDIA has published benchmark results for Nemotron 3.5 Super VL, including evaluations covering reasoning, coding, document understanding and agent tasks. These are useful reference points, but benchmark scores are not universal measures of business performance. Results depend on the test, model configuration, evaluation method and workload.

Enterprises should compare models using their own representative tasks, including accuracy, latency, cost, reliability and the rate at which human intervention is needed.

The Bigger Picture

Nemotron 3.5 Super VL reflects the transition from AI models that primarily answer questions toward systems designed to handle longer, multimodal workflows. Its combination of text, image and video understanding, a large context window and tool-use capabilities gives developers a foundation for building more specialised enterprise agents.

The release also illustrates the trade-offs involved in deploying advanced models. Open weights provide flexibility, but a 120-billion-parameter model still requires serious computing resources and careful implementation. Licensing, data availability, security and real-world performance will determine whether businesses can turn the model’s capabilities into reliable applications.

Looking Ahead

Developers will likely evaluate Nemotron 3.5 Super VL against other open-weight and proprietary models for tasks such as document processing, visual analysis, software development and enterprise automation. The most important question will be whether it performs reliably on practical workloads at an acceptable operating cost. The smaller Bolt models may also broaden the range of devices on which agentic applications can run.

For NVIDIA, the release expands its portfolio of customisable AI models while reinforcing its role in the infrastructure used to deploy them. For businesses, it offers another option to test, but not an automatic solution to agent development. The value will depend on the quality of the resulting applications, the safeguards around tool use and the ability to demonstrate measurable improvements in real workflows.

Frequently Asked Questions

1. What is NVIDIA Nemotron 3.5 Super VL?

It is an open-weight, multimodal AI model designed for enterprise agents. It can process text, images and video, reason through tasks, use tools and generate text responses.