Nemotron 3.5 Super VL's main payoff is bringing a high-capacity multimodal agent workload (reasoning, coding, tool-calling, long-context documents and video) into an operational GPU deployment profile by combining sparse MoE compute, Mamba convolutional layers, and an efficient NVFP4 quantization/distillation path.
Key Capabilities
- High-capacity multimodal backbone: hybrid Mamba-Transformer LatentMoE language trunk with a C-RADIO vision encoder and a vision-language connector — so it can jointly reason over text, images and sampled video frames in a single context window.
- Large but efficient footprint: 120B total parameters with ~12B active per forward pass and NVFP4 quantized weights — so you get strong capability while reducing memory and inference cost on supported NVIDIA hardware.
- Long-context and speculative decoding: supports up to 1M-token context and a shared Multi-Token Prediction (MTP) head for native speculative decoding — so long documents and throughput-sensitive deployments benefit from fewer verification steps and higher token throughput.
- Production-ready deployment recipes: tested runtimes include vLLM, SGLang and TensorRT-LLM with explicit configs for GB200, H100, H200 and DGX Spark — so teams can replicate performance on NVIDIA data-center hardware.
Who it's for and trade-offs
Great fit if you need a single model that handles coding agents, document/question answering over long contexts, and multimodal video/image understanding and you will deploy on NVIDIA GPU clusters (GB200/H100/H200/DGX). Look elsewhere if you require fast CPU-only inference, extremely small on-device models, or a simpler dense-only architecture — Nemotron Super VL relies on GPU-optimized runtimes, MoE routing, and quantization-aware distillation which increase system complexity and hardware requirements.
Where it fits
It sits between research-scale multimodal systems and production agent deployments: compared with dense VL models it reduces runtime cost via sparsity and NVFP4 quantization; compared with smaller VL models it offers stronger agentic and long-context capabilities but needs validated GPU serving infrastructure and adherence to the OpenMDW-1.1 governing terms.
