NVIDIA Launches Nemotron 3.5 Super VL for AI Agents

NVIDIA has released an open multimodal model called Nemotron 3.5 Super VL, built for AI agents that need to run nonstop and handle multi-step tasks. It uses a mixture-of-experts design with 120 billion total parameters and 12 billion active ones. NVIDIA says its throughput on a single GPU can reach up to 5 times that of similar open models.

The model combines reasoning, instruction following, and tool calling with text, image, document, and video understanding. It can help with code debugging, web searches, document analysis, and spotting cybersecurity threats. In tests, it scored 80.8% on the CharXiv chart reasoning benchmark, and 84% with code interpretation.

In one company case, a custom agent built by Scale AI is expected to cut inference costs by 97.5%. NVIDIA also launched two Nemotron 3.5 Bolt edge models. The model weights, training recipes, and datasets are now open on platforms like Hugging Face.