NVIDIA released Nemotron 3.5 Super VL in NVFP4 quantization. It is a 120B-parameter vision-language model with 12B active parameters. It supports up to 1M tokens of context. The model uses a hybrid Mamba-2, MoE and attention architecture. It combines a C-RADIO vision encoder with a vision-language connector. It is designed for reasoning, coding, tool use, document and video understanding. It is ready for commercial or non-commercial use. Deployment is supported on DGX Spark, H200, 2× H100 or GB200. It includes native Multi-Token Prediction for speculative decoding. The model is licensed under OpenMDW-1.1. It was published on Hugging Face on 15 October 2026.
What changed
The BF16 version required more memory and did not support NVFP4 quantization.
What it unlocks
Running a 120B-parameter vision-language model on a single DGX Spark, H200, 2× H100, or GB200 GPU.
- 120B total parameters, 12B active
- 1M token context length
- NVFP4 quantization
What you need to act on it
- NVIDIA DGX Spark, H200, 2× H100, or GB200 hardware
- Docker with NVIDIA GPU support
Sources