Nemotron 3.5 Super VL

NVIDIA Nemotron 3.5 Super VL is an open multimodal vision-language model engineered for complex, long-running agentic workflows and specialized enterprise tasks. Combining robust reasoning, instruction following, and tool use with native multimodal inputs, the model natively processes text, images, documents, GUI screenshots, and video streams. By unifying text and visual reasoning into a single system, it removes the need to route tasks between separate text and vision-language pipelines.

Architecture and Efficiency

The model features a 120B-parameter mixture-of-experts (MoE) architecture that activates 12B parameters per token to balance high model capacity with efficient inference. Its hybrid design integrates interleaved Mamba-2, Mixture-of-Experts, and attention layers alongside a C-RADIO vision encoder and a vision-language connector. It supports a context window of up to 1M tokens and incorporates Multi-Token Prediction (MTP) to enable native speculative decoding.

Capabilities and Customization

Nemotron 3.5 Super VL is optimized for advanced enterprise tasks including code generation, debugging, web research, document parsing, and complex multimedia investigation. Developers can further adapt and customize the model using NVIDIA NeMo with proprietary domain data, operational policies, and specialized toolsets. It is distributed in various precision formats, including full-precision BF16 weights and quantized NVFP4 options to match different infrastructure requirements.