AI agents are getting ambitious. They want to browse the web, read documents, call tools and finish jobs while people get on with their day. Nvidia’s latest Nemotron models are designed to give those agents a new set of options: bigger models for business systems, and smaller ones that can run locally on compatible PCs.
The lineup has three new models. Nemotron 3.5 Super VL is the heavyweight: a 120-billion-parameter model built for agents that need to understand text, images, charts and documents while carrying out multistep tasks. Nvidia says companies can customize it with their own data and run it on their own infrastructure, giving them more control over how an agent works and where information goes.
For PCs and workstations, Nvidia introduced two smaller models under the name Nemotron 3.5 Bolt. The 74-billion-parameter Bolt model is designed to fit within 24GB of memory on supported systems. It can power local agents that research, write, browse and debug code without sending those tasks to the cloud. The smaller Bolt E4B model is aimed at conversational agents, including voice assistants and interactive interfaces; Nvidia says it requires 4GB of memory and is due October 23.
That local setup could appeal to people who want an AI helper without routing every request through an online service. It also gives developers a way to experiment with agents on their own machines. There’s a catch, of course: “runs locally” depends on having compatible hardware, and a model’s ability to call tools doesn’t guarantee it will use them correctly. Agents still need sensible permissions and human oversight, especially when they can access files or take actions.
For businesses, the pitch is customization. A company could adapt Super VL to tasks such as reviewing financial documents, searching internal information or handling specialized workflows. Nvidia says organizations including EY, Ramp, Atlassian and CrowdStrike are working with or integrating the model. The company also claims Super VL can complete certain specialized tasks faster than competing models, but those results depend on the task and how the models are configured.
The models are “open weight,” meaning developers can access and adapt the model weights, with Nvidia also releasing datasets and training recipes. That offers more flexibility than a closed service, though it still takes technical skill and computing power to customize and operate a model well.
Final Byte
Nvidia’s new models give AI agents more places to run, from company systems to a desktop PC. The interesting test is whether they can make everyday work easier while staying predictable, private and firmly under their owner’s control. A helpful digital coworker is great; one that needs its own IT department is a different kind of project.
