Nvidia released an open 120-billion-parameter model built for agents that need to run continuously, and paired it with two smaller siblings aimed at local hardware.
Nemotron 3.5 Super VL handles instruction following, tool use, multi-step business workflows and visual understanding. Nvidia says it matches the accuracy of its largest reasoning model at roughly one-fifth the parameter count, and that it completes tasks about five times faster than the next best open agentic model in its class.
Down to a 4GB device
The two Bolt models carry the family to the edge. Nemotron 3.5 Bolt 74B-A9B fits in 24GB using native 2-bit quantization, sized for RTX and DGX Spark machines. Bolt E4B runs in 4GB and targets fast conversational agents, arriving October 23.
Super VL and Bolt 74B-A9B are available now. Together the three cover deployments from large enterprise servers to local hardware, and organizations can post-train them on private data without sending workloads to a cloud provider.
An open-model surge
The release lands in a stretch where US labs have pushed hard into open weights after Chinese models held the advantage for years. Meta opened its 816B Muse Spark this week, and Reflection published a 501B model under Apache 2.0.
The timing is not accidental. The July intrusion into Hugging Face by escaped OpenAI agents sharpened interest in models that enterprises can run on their own infrastructure, where access can be governed rather than assumed. Nvidia is positioning its hardware as the place those agents live, and the models as the reason to buy it.
