Nemotron 3.5 Super VL model can run locally on 128GB memory systems

The Nemotron 3.5 Super VL, a 120B vision-language Mixture-of-Experts (MoE) model released by NVIDIA, can be run locally on hardware with approximately 128GB of memory. While the model's headline checkpoint uses the NVFP4 4-bit format designed for Blackwell GPUs, community conversions including GGUF, MLX, FP8, and EXL3 have made it accessible on other hardware via Hugging Face.

Technical specifications for the model include 12B active parameters, a Mamba-2/Transformer hybrid architecture, and a C-RADIO vision encoder capable of processing text, images, and video up to two minutes in length. It supports up to 1M tokens of context and features native Multi-Token Prediction for speculative decoding.

Deployment options for local users include 128GB Mac systems or 96GB workstation cards using llama.cpp. Benchmarking tools and community data suggest that high memory capacity is a more critical factor for local execution than the specific GPU generation, noting that even high-end consumer cards like the RTX 5090 may be insufficient without adequate system or VRAM.