NVIDIA Details Nemotron 3.5 Super VL for Multimodal Agents

NVIDIA detailed Nemotron 3.5 Super VL in a company blog post. The open multimodal model uses a 120 billion parameter mixture of experts architecture and activates 12 billion parameters per token.

Nemotron 3.5 Super VL combines reasoning, instruction following, tool use, and support for text, images, documents, and video. NVIDIA says it provides up to five times higher throughput on one GPU than other open models in its size category.

Developers can customize the model with supervised fine tuning, low rank adaptation, and reinforcement learning through NVIDIA NeMo. NVIDIA also released datasets for tool use, finance, instruction following, long context tasks, and visual language training on Hugging Face.

The model supports LiLiCorr speculative decoding, multi token prediction, context reuse, and NVFP4 quantization. Deployment is supported through NVIDIA NIM and Dynamo.

NVIDIA says organizations including Scale AI, Ramp and AI21 have post-trained the model for specialized workflows. Scale AI expects a 40x lower inference cost for a vendor credit processing workload, while AI21 and Alice reported 41% higher accuracy on threat intelligence and vulnerability analysis.