Nemotron 3.5 Super VL 120B A12B BF16

Developers can customize or adapt the model to domain-specific data, and run full-precision inference and evaluation.

Availability

Hugging Face release; 10/15/2026

Pricing model

Not established

Deployment

4× H100 or 2× GB200

Privacy

Not established

Licence

OpenMDW License Agreement, version 1.1

• The model is released on Hugging Face as v1.0 - GA. ^[1]

  * [NVIDIA Nemotron 3 model family on Hugging Face](https://huggingface.co/collections/nvidia/nvidia-nemotron-v3)

• The model is designed for NVIDIA Blackwell or Hopper systems and integrates with PyTorch, Megatron-LM, and NeMo on Linux. ^[1]
• The model accepts text, images, and video and outputs text. ^{[1] [2] [3]}

  | **Inputs / Outputs** | Text, images, and video / text |

• The model supports context lengths up to 1M tokens. ^{[1] [2] [3] [4]}

  It has **120B total parameters** and **12B active parameters** and supports context lengths up to **1M tokens**.

• Native Multi-Token Prediction enables speculative decoding. ^{[1] [2] [3]}

  Multi-Token Prediction (MTP) enables native speculative decoding.

• BF16 weights are available on Hugging Face. ^[1]
• The documented BF16 deployment uses 4× H100 or 2× GB200. ^[2]

  The example below serves BF16 on **4× H100**.

• The model is intended for customization, domain adaptation, and full-precision evaluation. ^[2]

  The BF16 weights support customization, domain adaptation, and full-precision evaluation.