NVIDIA released Nemotron 3.5 Super VL on October 15 as an open-weight model built for long-running AI agents. It has 120 billion total parameters but activates 12 billion for each token, using a mixture-of-experts design that NVIDIA says balances model capacity with inference efficiency.
One model for text and visuals
The model combines text generation with image, document and video understanding, plus reasoning and tool use. NVIDIA's pitch is that an agent can read a PDF, inspect a chart, consult a webpage and act on the result without handing visual context between separate models. That could simplify workflows such as reviewing financial documents or applying company rules to scanned records, though developers still need to test whether the model extracts the right facts and follows those rules.
Organizations can post-train Super VL on their own data, tools and policies through NVIDIA's NeMo software. NVIDIA says it trained the model on about 26 trillion tokens, followed by roughly 3.8 million post-training samples and reinforcement learning across dozens of environments. The company is also releasing training recipes and datasets.
Speed claims come with boundaries
