NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
NVIDIA has released Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model designed for agentic workloads. Despite its large total parameter count, the model uses a mixture-of-experts (MoE) architecture that keeps only 12 billion parameters active during any single forward pass, reducing the compute required at inference time while preserving the capacity benefits of a much larger model.
A standout feature of the release is its support for a one-million-token context window, which is substantial even by current frontier model standards. That context length makes the model better suited for tasks involving long documents, extended multi-turn interactions, or complex visual inputs that require reasoning across many pieces of information at once. The model processes both text and images, fitting the multimodal agent use case NVIDIA is targeting.
On the hardware side, NVIDIA has tuned the model to run on a single DGX Spark unit using NVFP4 weights - a low-precision format that cuts memory bandwidth requirements without, in NVIDIA's framing, sacrificing meaningful accuracy. The ability to run a 120B-parameter-class model on one machine rather than a multi-node cluster is a practical consideration for organizations that want capable on-premise inference without large infrastructure investments. NVIDIA is releasing the model as open weights, which means developers can download and deploy it outside of NVIDIA's own cloud services.
Alongside the main Nemotron 3.5 Super VL model, NVIDIA is also releasing two Bolt edge models intended for deployment on more constrained hardware. The pairing follows a pattern seen across the industry, where a high-capacity model and smaller distilled or independently trained companions are released together to cover a wider range of deployment scenarios. Details on the Bolt models' parameter counts and supported hardware were not immediately specified, but their positioning as edge models suggests they are aimed at devices with significantly less memory and compute than a DGX system.
