NVIDIA put Nemotron 3.5 Super VL on Hugging Face. It is a 120 billion parameter MoE vision-language model. 12 billion parameters are active per token. It accepts vision and video. The context window is one million tokens.
What shipped
NVIDIA announced Nemotron 3.5 Super VL on October 15, 2026. 1
The public Hub name is NVIDIA-Nemotron-3.5-Super-VL-120B-A12B. It is a Mixture-of-Experts vision-language model. Total parameters are 120B. Active parameters per token are 12B. The router uses 512 experts with 22 active per token. Context window is 1M tokens.
License is OpenMDW-1.1. It permits commercial and non-commercial use. The license is permissive but non-standard.
Inputs are text, images, documents, and video. Video input is MP4 up to 2 minutes. Output is text.
Weights are public. NVFP4 is on Hugging Face. 2 BF16 is on Hugging Face. 5 ModelScope hosts a collection. 7 OpenRouter and NVIDIA NIM also list the model.
A technical report covers Super and Lightning. 6
What changed
This is a new generation. It is 3.5, not a patch of Nemotron 3 Super 120B. Nemotron 3 Super 120B was language-only. Super VL is vision-language.
Architecture is a Mamba-2 Transformer hybrid MoE with a C-RADIO vision encoder and Multi-Token Prediction. 3
The model is distilled from Nemotron 3 Ultra. Ultra is about 550B total with 55B active.
Training used about 26T tokens of pretraining plus continued pretraining including 10B Indic tokens. Post-training used about 3.8M samples plus RL.
Video input is supported. MP4 up to 2 minutes is allowed.
NVIDIA reports its own benchmarks. Those are vendor claims. Do not treat them as independent scores.
Independent coverage appeared on the same day. 4 8
Why it matters
12B active in a 120B MoE is the efficiency pitch. It offers more capacity than a 12B dense model with less work per token than a 120B dense model.
One checkpoint covers text, images, documents, and short video. Teams that previously used a text model plus a separate vision model can try one stack.
Context is 1M tokens.
Weights are on Hugging Face, ModelScope, NIM, and OpenRouter. Hosted paths exist if you do not want to run local iron.
License is OpenMDW-1.1, not Apache or MIT. Engineers should read it before commercial deploy.
It is a good fit for labs that want an open-weights 120B-class VL MoE with 12B active, 1M context, and video-in under a permissive non-standard license.
