Nvidia Nemotron 3.5 Super VL: Open Model for Always-On Enterprise Agents

What happened

Nvidia has released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters designed to power enterprise AI agents. According to reporting by The Decoder, the model is pitched as suitable for agents that run continuously rather than in short, intermittent sessions.

The release also includes two compact models, both named Bolt, which bring agent capability to local machines rather than relying solely on large-scale infrastructure. The Decoder reports that dozens of organisations are already fine-tuning Super VL for their own use.

The reporting notes one important qualification: getting the most out of Super VL means running it on Nvidia hardware. For organisations evaluating open models, that tie-in is central to the total cost and performance picture.

Why it matters

Open multimodal models that can handle text and visual inputs, and that are built for always-on agents, change what organisations can realistically automate. Agents that monitor systems, process documents, interpret images and respond to customers without constant human initiation move AI from a tool that people query into infrastructure that operates in the background. That shift affects how service is designed, staffed and governed.

The availability of compact models for local machines matters for the same reason. Some interactions carry privacy, latency or connectivity constraints that make cloud-only deployment impractical. Local agents give operations teams a second option, and the design decisions about where an agent runs will increasingly shape the experience customers and employees actually receive.

By the numbers

• 120 billion parameters: the size of the Nemotron 3.5 Super VL open multimodal model.
• Two models: the compact Bolt models released alongside Super VL for local machines.
• Dozens of organisations: the number already reported as fine-tuning Super VL.

The Renascence take

Most coverage of a model release fixates on parameter counts and benchmarks. The more useful question for an experience leader is where the agent sits in the service journey and who notices when it fails.

The headline is about nonstop agents, but the real design problem is continuity, not uptime. An agent that never sleeps still needs clear handoffs, visible limits and a way for customers to reach a person when the stakes rise. The behavioural principle underneath is that trust builds when people can predict what the system will do next. Operators should map which decisions agents can make alone, which need confirmation, and which must always reach a human, before they choose a model or a hardware stack.