Nemotron 3.5 Super VL: hardware and reasoning guide

NVIDIA researchers discuss Nemotron 3.5 Super VL, a vision-language model with 120 billion total parameters and 12 billion active parameters. The session explains its training and deployment choices for people considering a local LLM for document understanding or agent workloads.

For hardware, the speakers suggest DGX Spark with 128 GB of memory and certain Macs with more than 128 GB of unified memory. They do not provide an exact consumer GPU memory requirement. The host also describes Ollama CPU offloading as a way to use an RTX 4080 with system RAM, without giving performance measurements.

The team says it built on the Nemotron 3 text model, then added a vision model and projection through further training. Distillation from Ultra, a mid-training stage, filtered training traces and expanded reinforcement learning contributed to the reported improvements. Researchers also explain why mixture-of-experts components do not necessarily correspond to named specialties.

Vision topics include OCR, charts, long documents and video. Default video processing samples two frames per second, with adjustable sampling and resolution. The model is not a dedicated segmentation model. For self-hosted deployment, Venkat recommends starting with reasoning off, then trying medium and finally normal or high reasoning if answer quality falls short.