NVIDIA expands Nemotron family with new models

NVIDIA has today launched the Nemotron 3.5 Super VL, an open model with 120 billion parameters that provides the fastest task handling and highest accuracy for specialized, always-on agents. In addition to the Super VL, two new edge models have been introduced: the Nemotron 3.5 Bolt 74B-A9B 24GB, which is optimally tuned for intelligence and speed per GB for local workloads, and the Nemotron 3.5 Bolt E4B 4GB, aimed at rapid conversation agents. The Super VL and Bolt 74B-A9B are now available, while the Bolt E4B will be released on October 23.

These three open models cover the full spectrum of agent implementations, from large enterprise servers to edge devices and local hardware. Organizations looking to run agents on NVIDIA infrastructure now have an open, optimized model for their level, without having to choose between capability and deployability. In addition to the new models, the Nemotron family utilizes previous versions such as Lightning, Ultra, and Thread, which are aimed at various applications.

The Nemotron 3.5 Super VL is designed in collaboration with the Nemotron Coalition and offers leading accuracy for multimodal agentic tasks, such as following instructions and visual perception, approximately one-fifth of the parameters of NVIDIA's largest reasoning model. The model achieves up to five times faster throughput in its class, helping companies accelerate high-volume workflows. Organizations are able to further train the Super VL on their own data, enabling them to ensure the security and privacy of their information.

The Nemotron 3.5 Bolt 74B-A9B 24GB is specifically designed for local agentic tasks using NVIDIA RTX and DGX Spark, supporting the execution of tasks such as web research and code debugging without involving the cloud. The Bolt E4B 4GB focuses on physical and conversational agents within applications such as voice assistants and interactive characters. Across various industries, dozens of organizations have begun post-training the Super VL to improve accuracy for their specific tasks. Examples include company-specific applications such as scanning claims in the insurance sector and analyzing financial documents.

The openness and adaptability of the model provide maximum control over data and security. This is reinforced by the availability of open datasets and training recipes, allowing organizations to implement the model within their own infrastructure. NVIDIA facilitates these adaptations through its NeMo platform, offering a production-ready microservice package with regular updates for security and performance. The Nemotron 3.5 series is currently accessible via platforms like Hugging Face and ModelScope.

Read the full article from Nvidia.