Nvidia takes agents from data centers down to 4GB devices

On October 15, Nvidia introduced open Nemotron 3.5 models: Super VL with 120 billion parameters and two smaller Bolt variants for local devices. Super VL and Bolt 74B-A9B are available; the company says Bolt E4B is due on October 23.

Super VL targets enterprise workloads while Bolt runs locally

Super VL is designed for specialized agents working with text, images, documents, and video. Nvidia says its accuracy on instruction-following and tool-use tasks matches the larger Nemotron 3 Ultra despite having roughly one-fifth as many parameters.

Bolt 74B-A9B uses 2-bit quantization to fit within 24GB of memory and handle local tasks such as web browsing and code debugging. The smaller Bolt E4B, requiring 4GB, is aimed at conversational and physical agents.

Companies can tune the model to their own rules

The enterprise pitch is about more than raw performance. Organizations can post-train Super VL on their own data, tools, and workflows, then run it on their own infrastructure. Nvidia names EY, Ramp, Salesforce, and other companies among adopters, though their involvement alone does not show how well the model performs in routine production.

For IT teams, the lineup spans different memory and compute budgets. A local model may reduce the need to send sensitive tasks to the cloud, but data security and control still depend on device configuration, permissions, and updates.

Performance comparisons still rely largely on Nvidia’s claims

The company reports up to five times the speed of other models in its class and cites benchmark results. TechCrunch notes that launch comparisons are commonly part of model marketing. Without independent tests using the same tasks, hardware, and settings, those numbers cannot establish savings for a particular business.

Nor does “local” mean every PC can run a full agent workload. Nvidia specifies 24GB of memory and its own hardware for Bolt 74B-A9B, while the smaller E4B targets different, lighter use cases.

Real workstation deployments will decide the case

Watch for results from production deployments, not only post-training, and whether local agents can complete long tasks without frequent human intervention. For PC models, compatible hardware and practical speed and power use will matter in everyday work.

Enterprises will also need to assess auditability, access controls, and updates. Open weights offer more control, but shift operational responsibility to the team deploying them.

Lilith's verdict

Nvidia is extending the counter from the server room to the desktop. Companies can choose where an agent runs, but they still pick up the tab for supervising it.