NVIDIA's Nemotron 3.5 Super VL Pairs Vision With MoE

NVIDIA has published Nemotron 3.5 Super VL, a vision-language model built on a mixture-of-experts design, to Hugging Face. The release weighs in at 120 billion total parameters but activates only about 12 billion per token, a configuration NVIDIA labels "120B-A12B."

The model handles both image and text inputs and is described as multilingual, positioning it for document understanding, visual question answering, and other multimodal tasks where broad language coverage matters. It is distributed in BF16 format under a custom NVIDIA license rather than a standard permissive one, so teams should check the terms before deploying.

Why it matters

The appeal of a sparse MoE for vision-language work is cost. By routing each token through a small fraction of its experts, Nemotron 3.5 Super VL can offer the capacity of a large model while keeping the compute footprint closer to that of a 12B dense network.

• 120B total / 12B active: large capacity, lean per-token compute
• Vision + text: multimodal inputs for document and image reasoning
• Multilingual: aimed at use beyond English

As the first entry in the Nemotron 3.5 Super VL line, the model extends NVIDIA's push to supply open-weight building blocks for enterprise AI. Details on context length and benchmark results will be worth watching as the model card and community evaluations mature.