Nvidia has released three open-weight models aimed at running persistent AI agents across enterprise servers and personal computers. Nemotron 3.5 Super VL is a 120-billion-parameter model for larger deployments, while two smaller Bolt models are designed to run locally on PCs with RTX Spark chips and DGX Spark workstations.
The company says Super VL can approach the instruction-following and tool-use accuracy of its larger Nemotron 3 Ultra while requiring fewer parameters. Nvidia also claims strong task performance after specialized post-training in areas including document analysis, accounting, scientific work, insurance claims, and security. It names Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce among organizations using the larger model.
Those benchmark and adoption statements come from Nvidia and need independent evaluation across real workloads. Open weights give organizations more control over hosting and tuning than a closed API, but they do not remove infrastructure, safety, or data-governance work. The practical change is deployment choice: teams can select a server-scale multimodal model or a lighter local model while staying within Nvidia’s hardware and software ecosystem.