NVIDIA has introduced three new open models in the Nemotron 3.5 family, focused on running autonomous AI agents. According to the company, developers will be able to deploy them both on corporate infrastructure and on a standard personal computer. This was reported by RBC-Ukraine with reference to Interesting Engineering.
Central Release: The Multimodal Super VL
The key new addition is the Nemotron 3.5 Super VL multimodal model, featuring 120 billion parameters. It is designed for visual data analysis, document processing, and executing multi-step workflows using external tools.
Architecture: 120 Billion Parameters, 12 Billion Active
According to NVIDIA's official announcement, the model uses an architecture with a total of 120 billion parameters, of which 12 billion are active. This approach combines the capacity of a large neural network with a relatively low computational load per request, which is crucial for deployment on constrained hardware.
Speed and Accuracy According to the Company
According to NVIDIA, the model performs tasks five times faster than competitors in its class. At the same time, in terms of analysis and reasoning capabilities, the neural network approaches the company's largest models while utilizing only a fraction of their parameters.
Training and Fine-Tuning
The model was trained on 26 trillion tokens and supports fine-tuning via the NVIDIA NeMo framework, allowing companies to adapt it to their own tasks and data.
Developer Accessibility
Special attention is given to accessibility: the open Nemotron 3.5 family models can be run on enterprise infrastructure or even a standard PC, lowering the barrier to entry for teams lacking access to cloud clusters.
