Release2026-10-15
NVIDIA released Nemotron 3.5 Super VL in full-precision BF16. It is a vision-language model for reasoning, coding, tool use and document understanding. The model uses a Mamba-2–Transformer hybrid Mixture-of-Experts architecture. It has 120B total parameters and 12B active parameters. It supports up to 1M tokens of context. It is licensed under OpenMDW-1.1. It is ready for commercial or non-commercial use. It is deployed on 4× H100 or 2× GB200. It supports native speculative decoding via Multi-Token Prediction. It is available globally on Hugging Face.
What it unlocks
Running a vision-language model with 1M token context on images, video and text for coding, reasoning and tool use.
- 120B total parameters
- 12B active parameters
- 1M token context
- 4× H100 or 2× GB200 deployment
What you need to act on it
- 4× H100 or 2× GB200 GPUs
- Docker with NVIDIA GPU support
Sources