NVIDIA Releases Full-Weight Nemotron 3.5 Super VL on HuggingFace

On October 12, 2026, NVIDIA published the NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 model on HuggingFace with the image-text-to-text task. This is a full-weight release of a vision-language model for reasoning, code, document work and video understanding.

What it means

The model has 124.5 billion parameters, of which 12 billion are active — a MoE architecture with hybrid Mamba-2 and attention layers. The total weights in the repository are 231.8 GB, the largest file is 3.7 GB. It runs via transformers, the weights are stored in safetensors. The OpenMDW-1.1 license permits free use, modification and distribution, including commercial use. When distributing, you must include the license and retain the attribution notices. Rights terminate if you file a patent or copyright infringement lawsuit related to the model's materials. There are no restrictions on the outputs. For local running, the platform has 12 quantized builds: NVFP4, GGUF, FP8, MLX and EXL. The smallest working one is 64.6 GB at 4 bit. The full-weight version requires 4x H100 or 2x GB200, context — up to 1M tokens. In the same days, NVIDIA released NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16. You can follow the appearance of new models and builds in the CPU3D news feed.