Nvidia has released Nemotron 3.5 Super VL, a 120-billion-parameter open multimodal model built to run enterprise AI agents continuously. Developed alongside the Nemotron Coalition, the model uses a Mixture-of-Experts architecture with 12 billion active parameters per token to handle text, images, documents, and videos. Alongside the flagship model, Nvidia introduced two smaller Bolt variants, including a 24GB version for local machines and a forthcoming 4GB version for voice assistants.
Nvidia distilled Super VL from its Nemotron 3 Ultra reasoning model, achieving comparable accuracy with only one-fifth of the parameters. The model runs on Nvidia's NVFP4 4-bit precision format, and features a built-in draft component and LiLiCorr method to accelerate text generation. Nvidia is publishing the model weights, datasets, and training recipes to help enterprises fine-tune the technology on their own servers for tasks like fraud detection and credit scoring.
This release matters because it gives enterprises high-performance, customizable open-weights models optimized to run continuous agent tasks while driving hardware sales for Nvidia.