Nemotron 3.5 Super VL 120B A12B NVFP4
NVIDIA Nemotron 3.5 Super VL is a multimodal Mixture-of-Experts model with 120B total parameters, 12B active per token, and up to one million tokens of context. This NVFP4 version is built for visual reasoning, document understanding, and agentic workflows across images, video, and text.
Model
NVIDIA Nemotron 3.5 Super VL is an open-weight multimodal reasoning model designed for sophisticated agentic workflows involving text, images, documents, and video. It supports document and chart understanding, visual reasoning over screenshots, video question answering, and tool-enabled applications.
Built on a hybrid Mamba-Transformer Mixture-of-Experts architecture with NVIDIA's C-RADIOv4 vision encoder, the model combines 120B total parameters with just 12B active parameters per token, balancing reasoning capability with inference efficiency.
The model supports up to 1 million tokens of context, configurable reasoning, and built-in Multi-Token Prediction. This NVFP4 version uses lower-precision language-model weights, while the vision encoder and projector remain in BF16.
See NVIDIA's announcement for full technical details.
Dedicated Endpoint
DocsA Dedicated Endpoint is your own deployment of Nemotron 3.5 Super VL 120B A12B NVFP4, on an inference stack that Modal tunes and autoscales. Bring fine-tuned weights if you have them. Compute is billed by the second, and only while it runs.
modal endpoint create --model nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4