Nvidia expands the Nemotron LLM family with highly efficient models for agent systems

Nvidia Corp. today announced the launch of customizable Nemotron 3.5 Super LV large language models for always-on artificial intelligence agents and edge-focused 3.5 Bolt for local systems.

The company said the three new models will cover the full spectrum of agent deployments, from large enterprise servers to edge devices and on-premises hardware.

Nemotron 3.5 Super VL is a 120 billion parameter open model developed in collaboration with the Nemotron Coalition, a group of model builders and AI labs advancing frontier-level open models designed for high-precision multimodal agent tasks. It is able to follow instructions, use tools, perform multi-step business processes and understand visually. Nvidia said it had comparable performance to the company’s largest reasoning models at just a fifth of the size.

Thanks to this size and a new architecture, the Super VL can produce results almost five times faster than other models in the same class.

Unlike closed-frontier models, Nvidia makes the Super VL weights openly available, meaning companies can subsequently train them on their own data and deploy them on their own infrastructure. This means they can put it behind their firewall and keep it within their domain, maintaining sovereignty, privacy and security.

To develop Super VL, Nvidia trained the model with open data and open methods, using approximately 26 trillion tokens during pretraining and further pretraining. This also included 3.8 million post-training examples and reinforcement learning in 100 open environments to fine-tune the model.

Nvidia also announced that it is releasing open data sets on Hugging Face, which include agent tool usage, finance, instruction following, long context, security, fairness, and vision language understanding.

Bolt 74B-A9B 24GB is a smaller model designed for local execution of agent tasks and is based on Super VL. It uses a hybrid expert mixture model that activates 9 billion of its 74 billion parameters simultaneously and uses 2-bit quantization to fit into 24 gigabytes of memory. This allows it to conveniently run on an enterprise workstation or PC GPUs, allowing users to delegate everyday projects, including writing and debugging code, without sending workloads to the cloud.

Later this month, Nvidia is also launching an even smaller model: Bolt E4B 4GB. This tiny model is designed for edge use on GeForce RTX, Jetson and RTX/GTX Spark. This enables physical AI and conversational agents to work in devices such as wearables, home appliances, point-of-service, car dashboards, device assistants and other embedded interfaces.

Nvidia noted that it could power interactive characters, conversational interfaces and intelligent user interfaces for localization or personalization.

Image: Nvidia