NVIDIA Nemotron 3.5 Super VL

Vision-language reasoning model trained by NVIDIA. Pairs a hybrid Mixture-of-Experts language backbone (interleaved Mamba-2 and MoE layers, with select Attention layers) with a C-RADIO v4-H vision encoder. Contains 120B total parameters with only 12B active at a time, and includes a built-in Multi-Token Prediction (MTP) head for faster generation.

Accepts image and text input for document intelligence, GUI understanding, and visual Q&A. Features a reasoning toggle to enable or disable intermediate reasoning traces, and an optional medium reasoning effort mode. Includes native agentic capabilities for tool use and structured output. Supports English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, and Chinese.

Supports a context length of 1M tokens.