NVIDIA released Nemotron 3.5 Super VL, a new open model that combines reasoning and tool use with text, image, document and video understanding, aimed at long-running AI agents. It has 120B total parameters with 12B active, so only 10% of the weights are used for any given token. NVIDIA AI announced it on X on October 15.
From NVIDIA AI's announcement.
What it is
The release puts several capabilities in one model. Reasoning and tool use are what an agent needs to plan and act over many steps, while text, image, document and video understanding let the same model work with different kinds of input.
NVIDIA's demo runs inside Hermes Agent, an agent interface with Chat, Sessions, Files, Models, Logs, Cron and Skills in its sidebar. In a new session, the prompt is 'give me a morning brief'. The model shows a short thinking step of about 52 tokens, concluding that the user wants a morning brief and that it should load the matching skill. It then makes two tool calls: it views a skill whose name begins with 'morning-brief', and it runs a Python command in the terminal. Two more short thinking steps of about 47 and 36 tokens follow, with running totals of about 54 and 66 tokens.
Context
Nemotron 3.5 Super VL follows Nemotron 3.5 Lightning, the August release with 30B total and about 3B active parameters, which has the same 10% active ratio at a quarter of the size. NVIDIA's other recent open releases include a teacher model behind Nemotron 3 Ultra, its 550B flagship.