Nvidia expands Nemotron LLM family with hyper-efficient models for agentic systems

Nvidia today widened its Nemotron family with three new models aimed at agentic systems, and the spread is the point. One is for large enterprise setups, one is for local machines, and one is built for edge hardware. The company is clearly trying to cover every place an AI agent might live, instead of forcing everything through the cloud.

The headliner is Nemotron 3.5 Super VL, an open 120-billion-parameter model built with the Nemotron Coalition, a group of model builders and AI labs focused on open frontier models. Nvidia says it is designed for high-accuracy multimodal agent work, including instruction following, tool use, multistep business workflows and visual understanding. The company also says the model has comparable power to its largest reasoning models while being one-fifth the size, and that the new architecture makes it nearly five times faster than other models in the same class.

Open weights are the real strategic move here. Nvidia says organizations can post-train Super VL on their own data and run it on their own infrastructure, which keeps it behind the firewall and inside their own domain. The company also says it trained the model with open data and open methods using about 26 trillion tokens across pretraining and continued pretraining, plus 3.8 million post-training samples and reinforcement learning across 100 open environments. Alongside that, Nvidia is putting out open datasets on Hugging Face covering agentic tool use, finance, instruction following, long context, safety, fairness and vision-language understanding.

The local option is Bolt 74B-A9B 24GB, a smaller model based on Super VL. It uses a mixture-of-experts setup that activates 9 billion of its 74 billion parameters at a time, plus two-bit quantization to fit within 24 gigabytes of memory. Nvidia says that makes it comfortable on an enterprise workstation or PC graphics cards, with uses like writing and debugging code without sending the work to the cloud.

Later this month, Nvidia plans to release an even smaller model, Bolt E4B 4GB, for edge deployment on GeForce RTX, Jetson and RTX/GTX Spark. That one is aimed at wearables, household devices, point-of-service systems, car dashboards, device assistants and other embedded interfaces. In other words, Nvidia wants agents everywhere: in the datacenter, on the desktop and tucked inside the things people actually touch.