Nvidia launches new open models to run agents in the enterprise and on PCs

Nvidia launched the Nexus Super VL, a 120-billion-parameter model built for enterprise agents. It is designed for instruction following and tool use, including document intelligence, accounting, scientific work, medical claims, and security tasks. Nvidia says it can match the accuracy of its larger Nexus Ultra on these tasks.

The company also introduced two lighter Bolt models. Nexus Bolt 74B-A9B 24GB is meant for local agentic work, including multistep reasoning, tool calling, and web browsing. The smaller Bolt E4B focuses on conversational agents.

Super VL and Bolt 74B-A9B are available now. Bolt E4B is expected later this month. Together, the models cover different deployment needs, from enterprise data centers to AI PCs and workstations using Nvidia chips. Nvidia says organizations can choose models based on capability and deployability.

An open-weight AI model is released with its trained weights available for others to download, run, and often adapt. That gives organizations more control over where the model operates and how it is customized. The article describes Nvidia’s Super VL as open, allowing organizations to post-train it on their own data and deploy it on their own infrastructure.

A closed model is controlled by its provider. Users typically access it through a hosted service or approved interface, rather than receiving the model weights for independent deployment. The article contrasts Nvidia’s models with proprietary frontier labs such as Paradox and Axiom.

The distinction matters for control, privacy, and customization. Nvidia says enterprises can own what they build with Super VL. Open weights still require suitable hardware, technical expertise, and responsible management, so they do not automatically make deployment simple or inexpensive.

Nvidia’s Nexus Super VL has 120 billion parameters. Nexus Bolt 74B-A9B 24GB has 74 billion parameters, while the article names the smaller model as Bolt E4B without stating its parameter count. Super VL is large enough for enterprise agent workloads, while Bolt models target more local, deployable use.

Parameters are adjustable values learned during training. They help a model represent language, visual information, instructions, and patterns for reasoning. More parameters can increase a model’s capacity, but size alone does not guarantee better results. Nvidia says Super VL can deliver accuracy similar to the larger Nexus Ultra on certain tasks.

Model size also affects hardware needs. Larger models generally require more memory and computing power to run efficiently. That is why Nvidia positions Super VL for enterprise infrastructure and Bolt models for AI PCs and workstations, including the 24GB Bolt designation.

A basic chatbot mainly responds to the text a user sends. An AI agent can pursue a broader task by deciding what steps are needed, using available tools, and checking or combining results. The article connects agentic models with tool use, web browsing, document work, accounting, and other specialized tasks.

For example, an agent could receive a request involving several documents, use a tool to inspect them, search for additional information, and then produce an organized result. Nvidia says Bolt 74B-A9B is intended for multistep reasoning, tool calling, and web browsing. Those abilities let the model act as a coordinator, not only a conversational responder.

Agents still depend on the tools, permissions, data, and instructions provided to them. The article presents them as useful for enterprise workflows and everyday projects. Nvidia also wants local agents to handle tasks without sending workloads to the cloud, particularly on compatible PCs.

When an enterprise runs a model on its own infrastructure, its workloads do not have to be sent to a cloud AI provider. Nvidia says organizations can post-train Super VL on their own data, deploy it in their own data centers, and own what they build. That can give them greater control over sensitive business information and the model’s behavior.

A company might adapt Super VL for document intelligence, accounting, insurance claims, or security work using internal data. The model could then operate within the company’s Nvidia infrastructure instead of depending entirely on an external hosted service. Nvidia says companies including EY, Ramp, Salesforce, and others are using Super VL.

The trade-off is responsibility. The enterprise must provide compatible hardware, operate the system, protect the data, and manage updates and performance. Nvidia is targeting this demand as proprietary AI labs seek other chips and companies increasingly want open models in their own data centers.

Consumers may prefer local assistants because they can keep everyday workloads on their own PCs. Nvidia specifically points to privacy concerns around assistants accessing family, health, and financial information. Its Bolt models are designed to run on AI PCs using the Nvidia RTX Spark chip and on DGX Spark workstations.

A local assistant could handle conversation, multistep projects, tool calls, or browsing directly on the computer. That means the workload need not be sent to a cloud service. Nvidia presents this as a way for users to delegate everyday projects while retaining more control over where their information is processed.

Local operation has trade-offs. The computer must have enough memory and processing capability, and the user or organization must manage the software. Local models may also be smaller than the largest cloud systems. Nvidia is using these models to support a growing, but still nascent, market for consumer AI assistants.

Large language models require enormous amounts of computation during training and use. GPUs and other specialized AI chips accelerate many calculations in parallel and provide the memory and throughput needed to process model parameters. Nvidia’s new models are tied to its infrastructure, including RTX Spark chips for AI PCs and DGX Spark workstations.

The hardware supports different deployment choices. Large models such as Super VL are aimed at enterprise systems, while smaller Bolt models are designed for local machines. Specialized systems help models perform tasks such as reasoning, tool calling, web browsing, and conversational assistance at usable speeds.

Nvidia is not the only hardware option. The article says proprietary AI labs are looking to systems from AWS, including Trainium, and Cerebras, while also developing home-grown chips. These efforts reflect a desire for more choices as AI workloads grow and companies seek ways to power their models beyond Nvidia infrastructure.